Steven E. Pav steven@cerebellumcapital - R in...
Transcript of Steven E. Pav steven@cerebellumcapital - R in...
Portfolio inference with this one weird trick
Steven E Pavstevencerebellumcapitalcom
Cerebellum Capital
May 16 2014
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 1 36
One Weird Trick
A weird
Consider p-vector of asset returns x Let
micro = E [x] Σ = Var (x)
Prepend a lsquo1rsquo to the vector x =[1 xgt
]gt
The second moment of x contains the first two moments of x
Θ = E[xxgt
]=
[1 microgt
micro Σ + micromicrogt
]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 2 36
One Weird Trick
trick
then Θminus1 =
[1 + microgtΣminus1micro minusmicrogtΣminus1
minusΣminus1micro Σminus1
]
=
1 + ζ2lowast minusνlowastgt
minusνlowast Σminus1
νlowast is the Markowitz portfolio
ζlowast is the Sharpe ratio of νlowast (cf Hotellingrsquos T 2)
Σminus1 is the lsquoprecision matrixrsquo
The portfolio is lsquooptimalrsquo solving eg Royrsquos problem [17]
νlowast prop argmaxννgtΣνleR2
νgtmicrominus r0radicνgtΣν
ie ldquomaximize Sharpe with a bound on riskrdquoSteven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 3 36
One Weird Trick
But is it useful
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 4 36
Portfolio Inference
Sample estimator
Since Θ = E[xxgt
]the simple estimator is unbiased
Θ =1
n
sum1leilen
xi xigt =
[1 microgt
micro Σ + micromicrogt
]
The inverse contains the sample estimates
Θminus1 =
1 + ζ2lowast minusνlowastgt
minusνlowast Σminus1
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 5 36
Portfolio Inference
Asymptotics I
By the Central Limit Theorem
radicn(
vech(
Θ)minus vech (Θ)
) N (0Ω)
where Ω = Var(vech
(xxgt
))
We can estimate Ω from the sample call it ΩItrsquos just sample covariance of vech
(xi xi
gt) for 1 le i le n
Use the delta method
radicn(
vech(
Θminus1)minus vech
(Θminus1
)) N
(0UΩUgt
)
Here U is some lsquouglyrsquo derivative depending on Θ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 6 36
Portfolio Inference
Asymptotics II
Ignoring details about symmetry etc the derivative is [7 12]
dXminus1
dX= minus
(Xminusgt otimes Xminus1
)
(This generalizes the scalar derivative)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 7 36
Portfolio Inference
I can make a hat or a brooch or a pterodactyl
Θminus1 =
1 + ζ2lowast minusνlowastgt
minusνlowast Σminus1
What is the use for Var
(vech
(Θminus1
))
Perform inference on elements of νlowast via Wald statistic(Compare elements of νlowast to their standard errors)
Perform inference on the maximal Sharpe ratio ζlowast
Equivalently Hotellingrsquos T 2 test (tests hypothesis micro is all zeros)
Portfolio shrinkage
Estimate the covariance of νlowast and Σminus1 (Attribute portfolio error toreturns or covariance) [5]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 8 36
Portfolio Inference
Implementation trust but verify
require(MarkowitzR)
setseed(2014)
X lt- matrix(rnorm(1000 5) ncol = 5) toy data
ism lt- MarkowitzRmp_vcov(X)
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
print(t(walds(ism))) Wald stats
X1 X2 X3 X4 X5
Intercept 089 -022 -16 -24 -049
cf Britten-Jones httpjstororgstable2697722
y lt- rep(1 dim(X)[1])
print(t(summary(lm(y ~ X - 1))$coefficients[ 3]))
X1 X2 X3 X4 X5
[1] 089 -022 -16 -25 -048
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 9 36
Portfolio Inference
Game over
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 10 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Extensions
Hedged portfolios I
Hedging the goal
Returns which are statistically independent from some random variables
Hedging a more realistic goal
A portfolio with zero covariance to some random variables
Hedging an achievable goal
A portfolio with zero sample covariance to some other portfolios oftradeable assets(eg you may have to hold some Mkt to hedge out the Mkt)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 12 36
Extensions
Hedged portfolios II
maxνGΣν=0νgtΣνleR2
νgtmicrominus r0radicνgtΣν
where G is a pg times p matrix of rank pg
Rows of G define portfolios against which we have 0 covariance
Typically G consists of some rows of identity matrix
ie ldquoMaximize Sharpe ratio with risk bound and zero covariance to someother portfoliosrdquoSolved by cνGlowast with c to satisfy risk bound and
νGlowast =(
Σminus1microminus Ggt(GΣGgt
)minus1Gmicro)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 13 36
Extensions
Hedged portfolios III
Use the weird trick Let G =
[1 00 G
] then
Θminus1 minus Ggt(
GΘGgt)minus1
G = microgtΣminus1microminus microgtGgt(GΣGgt
)minus1Gmicro minusνGlowast
gt
minusνGlowast Σminus1 minus Ggt(GΣGgt
)minus1G
minusνGlowast is the optimal hedged portfolio
UL corner is squared Sharpe ratio of νGlowastAlso used for portfolio spanning [16 8 10 11]
LR corner is loss of precision
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 14 36
Extensions
Hedged portfolios IV
Delta method gives the asymptotic distribution
radicn(
vech(
∆GΘminus1)minus vech
(∆GΘminus1
)) N
(0UΩUgt
)
with more ugly derivatives
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 15 36
Extensions
Hedged portfolios V
Download the Fama-French 3 factor + Momentum monthly data(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market compute (unhedged) Markowitzportfolio and Wald statistics
wstats lt- rbind(doboth(ff4xts[ 14]) wtrickws(ff4xts[
14] vcovfunc = sandwichvcovHAC))
rownames(wstats)[3] lt- c(weird trick w HAC)
xtable(wstats)
Mkt SMB HML UMD
Britten Jones t-stat 628 072 499 820weird trick Wald stat 537 077 447 603weird trick w HAC 510 077 392 553
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 16 36
Extensions
Hedged portfolios VI
Now hedge out Mkt
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
Gmat lt- matrix(diag(1 4)[1 ] ncol = 4)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 14] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD
Intercept 269 077 447 603
And compute the spanning Wald statistic
efstat lt- function(ism) ism$mu[1]sqrt(ism$Ohat[1 1])
print(efstat(asymv))
[1] 38
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 17 36
Extensions
Hedged portfolios VII
Now hedge out Mkt and RF
hedge out RFR too
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 15] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD RF
Intercept 071 205 227 343 -130
And the spanning statistic
print(efstat(asymv))
[1] 21
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 18 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
One Weird Trick
A weird
Consider p-vector of asset returns x Let
micro = E [x] Σ = Var (x)
Prepend a lsquo1rsquo to the vector x =[1 xgt
]gt
The second moment of x contains the first two moments of x
Θ = E[xxgt
]=
[1 microgt
micro Σ + micromicrogt
]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 2 36
One Weird Trick
trick
then Θminus1 =
[1 + microgtΣminus1micro minusmicrogtΣminus1
minusΣminus1micro Σminus1
]
=
1 + ζ2lowast minusνlowastgt
minusνlowast Σminus1
νlowast is the Markowitz portfolio
ζlowast is the Sharpe ratio of νlowast (cf Hotellingrsquos T 2)
Σminus1 is the lsquoprecision matrixrsquo
The portfolio is lsquooptimalrsquo solving eg Royrsquos problem [17]
νlowast prop argmaxννgtΣνleR2
νgtmicrominus r0radicνgtΣν
ie ldquomaximize Sharpe with a bound on riskrdquoSteven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 3 36
One Weird Trick
But is it useful
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 4 36
Portfolio Inference
Sample estimator
Since Θ = E[xxgt
]the simple estimator is unbiased
Θ =1
n
sum1leilen
xi xigt =
[1 microgt
micro Σ + micromicrogt
]
The inverse contains the sample estimates
Θminus1 =
1 + ζ2lowast minusνlowastgt
minusνlowast Σminus1
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 5 36
Portfolio Inference
Asymptotics I
By the Central Limit Theorem
radicn(
vech(
Θ)minus vech (Θ)
) N (0Ω)
where Ω = Var(vech
(xxgt
))
We can estimate Ω from the sample call it ΩItrsquos just sample covariance of vech
(xi xi
gt) for 1 le i le n
Use the delta method
radicn(
vech(
Θminus1)minus vech
(Θminus1
)) N
(0UΩUgt
)
Here U is some lsquouglyrsquo derivative depending on Θ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 6 36
Portfolio Inference
Asymptotics II
Ignoring details about symmetry etc the derivative is [7 12]
dXminus1
dX= minus
(Xminusgt otimes Xminus1
)
(This generalizes the scalar derivative)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 7 36
Portfolio Inference
I can make a hat or a brooch or a pterodactyl
Θminus1 =
1 + ζ2lowast minusνlowastgt
minusνlowast Σminus1
What is the use for Var
(vech
(Θminus1
))
Perform inference on elements of νlowast via Wald statistic(Compare elements of νlowast to their standard errors)
Perform inference on the maximal Sharpe ratio ζlowast
Equivalently Hotellingrsquos T 2 test (tests hypothesis micro is all zeros)
Portfolio shrinkage
Estimate the covariance of νlowast and Σminus1 (Attribute portfolio error toreturns or covariance) [5]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 8 36
Portfolio Inference
Implementation trust but verify
require(MarkowitzR)
setseed(2014)
X lt- matrix(rnorm(1000 5) ncol = 5) toy data
ism lt- MarkowitzRmp_vcov(X)
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
print(t(walds(ism))) Wald stats
X1 X2 X3 X4 X5
Intercept 089 -022 -16 -24 -049
cf Britten-Jones httpjstororgstable2697722
y lt- rep(1 dim(X)[1])
print(t(summary(lm(y ~ X - 1))$coefficients[ 3]))
X1 X2 X3 X4 X5
[1] 089 -022 -16 -25 -048
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 9 36
Portfolio Inference
Game over
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 10 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Extensions
Hedged portfolios I
Hedging the goal
Returns which are statistically independent from some random variables
Hedging a more realistic goal
A portfolio with zero covariance to some random variables
Hedging an achievable goal
A portfolio with zero sample covariance to some other portfolios oftradeable assets(eg you may have to hold some Mkt to hedge out the Mkt)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 12 36
Extensions
Hedged portfolios II
maxνGΣν=0νgtΣνleR2
νgtmicrominus r0radicνgtΣν
where G is a pg times p matrix of rank pg
Rows of G define portfolios against which we have 0 covariance
Typically G consists of some rows of identity matrix
ie ldquoMaximize Sharpe ratio with risk bound and zero covariance to someother portfoliosrdquoSolved by cνGlowast with c to satisfy risk bound and
νGlowast =(
Σminus1microminus Ggt(GΣGgt
)minus1Gmicro)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 13 36
Extensions
Hedged portfolios III
Use the weird trick Let G =
[1 00 G
] then
Θminus1 minus Ggt(
GΘGgt)minus1
G = microgtΣminus1microminus microgtGgt(GΣGgt
)minus1Gmicro minusνGlowast
gt
minusνGlowast Σminus1 minus Ggt(GΣGgt
)minus1G
minusνGlowast is the optimal hedged portfolio
UL corner is squared Sharpe ratio of νGlowastAlso used for portfolio spanning [16 8 10 11]
LR corner is loss of precision
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 14 36
Extensions
Hedged portfolios IV
Delta method gives the asymptotic distribution
radicn(
vech(
∆GΘminus1)minus vech
(∆GΘminus1
)) N
(0UΩUgt
)
with more ugly derivatives
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 15 36
Extensions
Hedged portfolios V
Download the Fama-French 3 factor + Momentum monthly data(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market compute (unhedged) Markowitzportfolio and Wald statistics
wstats lt- rbind(doboth(ff4xts[ 14]) wtrickws(ff4xts[
14] vcovfunc = sandwichvcovHAC))
rownames(wstats)[3] lt- c(weird trick w HAC)
xtable(wstats)
Mkt SMB HML UMD
Britten Jones t-stat 628 072 499 820weird trick Wald stat 537 077 447 603weird trick w HAC 510 077 392 553
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 16 36
Extensions
Hedged portfolios VI
Now hedge out Mkt
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
Gmat lt- matrix(diag(1 4)[1 ] ncol = 4)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 14] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD
Intercept 269 077 447 603
And compute the spanning Wald statistic
efstat lt- function(ism) ism$mu[1]sqrt(ism$Ohat[1 1])
print(efstat(asymv))
[1] 38
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 17 36
Extensions
Hedged portfolios VII
Now hedge out Mkt and RF
hedge out RFR too
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 15] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD RF
Intercept 071 205 227 343 -130
And the spanning statistic
print(efstat(asymv))
[1] 21
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 18 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
One Weird Trick
trick
then Θminus1 =
[1 + microgtΣminus1micro minusmicrogtΣminus1
minusΣminus1micro Σminus1
]
=
1 + ζ2lowast minusνlowastgt
minusνlowast Σminus1
νlowast is the Markowitz portfolio
ζlowast is the Sharpe ratio of νlowast (cf Hotellingrsquos T 2)
Σminus1 is the lsquoprecision matrixrsquo
The portfolio is lsquooptimalrsquo solving eg Royrsquos problem [17]
νlowast prop argmaxννgtΣνleR2
νgtmicrominus r0radicνgtΣν
ie ldquomaximize Sharpe with a bound on riskrdquoSteven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 3 36
One Weird Trick
But is it useful
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 4 36
Portfolio Inference
Sample estimator
Since Θ = E[xxgt
]the simple estimator is unbiased
Θ =1
n
sum1leilen
xi xigt =
[1 microgt
micro Σ + micromicrogt
]
The inverse contains the sample estimates
Θminus1 =
1 + ζ2lowast minusνlowastgt
minusνlowast Σminus1
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 5 36
Portfolio Inference
Asymptotics I
By the Central Limit Theorem
radicn(
vech(
Θ)minus vech (Θ)
) N (0Ω)
where Ω = Var(vech
(xxgt
))
We can estimate Ω from the sample call it ΩItrsquos just sample covariance of vech
(xi xi
gt) for 1 le i le n
Use the delta method
radicn(
vech(
Θminus1)minus vech
(Θminus1
)) N
(0UΩUgt
)
Here U is some lsquouglyrsquo derivative depending on Θ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 6 36
Portfolio Inference
Asymptotics II
Ignoring details about symmetry etc the derivative is [7 12]
dXminus1
dX= minus
(Xminusgt otimes Xminus1
)
(This generalizes the scalar derivative)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 7 36
Portfolio Inference
I can make a hat or a brooch or a pterodactyl
Θminus1 =
1 + ζ2lowast minusνlowastgt
minusνlowast Σminus1
What is the use for Var
(vech
(Θminus1
))
Perform inference on elements of νlowast via Wald statistic(Compare elements of νlowast to their standard errors)
Perform inference on the maximal Sharpe ratio ζlowast
Equivalently Hotellingrsquos T 2 test (tests hypothesis micro is all zeros)
Portfolio shrinkage
Estimate the covariance of νlowast and Σminus1 (Attribute portfolio error toreturns or covariance) [5]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 8 36
Portfolio Inference
Implementation trust but verify
require(MarkowitzR)
setseed(2014)
X lt- matrix(rnorm(1000 5) ncol = 5) toy data
ism lt- MarkowitzRmp_vcov(X)
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
print(t(walds(ism))) Wald stats
X1 X2 X3 X4 X5
Intercept 089 -022 -16 -24 -049
cf Britten-Jones httpjstororgstable2697722
y lt- rep(1 dim(X)[1])
print(t(summary(lm(y ~ X - 1))$coefficients[ 3]))
X1 X2 X3 X4 X5
[1] 089 -022 -16 -25 -048
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 9 36
Portfolio Inference
Game over
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 10 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Extensions
Hedged portfolios I
Hedging the goal
Returns which are statistically independent from some random variables
Hedging a more realistic goal
A portfolio with zero covariance to some random variables
Hedging an achievable goal
A portfolio with zero sample covariance to some other portfolios oftradeable assets(eg you may have to hold some Mkt to hedge out the Mkt)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 12 36
Extensions
Hedged portfolios II
maxνGΣν=0νgtΣνleR2
νgtmicrominus r0radicνgtΣν
where G is a pg times p matrix of rank pg
Rows of G define portfolios against which we have 0 covariance
Typically G consists of some rows of identity matrix
ie ldquoMaximize Sharpe ratio with risk bound and zero covariance to someother portfoliosrdquoSolved by cνGlowast with c to satisfy risk bound and
νGlowast =(
Σminus1microminus Ggt(GΣGgt
)minus1Gmicro)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 13 36
Extensions
Hedged portfolios III
Use the weird trick Let G =
[1 00 G
] then
Θminus1 minus Ggt(
GΘGgt)minus1
G = microgtΣminus1microminus microgtGgt(GΣGgt
)minus1Gmicro minusνGlowast
gt
minusνGlowast Σminus1 minus Ggt(GΣGgt
)minus1G
minusνGlowast is the optimal hedged portfolio
UL corner is squared Sharpe ratio of νGlowastAlso used for portfolio spanning [16 8 10 11]
LR corner is loss of precision
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 14 36
Extensions
Hedged portfolios IV
Delta method gives the asymptotic distribution
radicn(
vech(
∆GΘminus1)minus vech
(∆GΘminus1
)) N
(0UΩUgt
)
with more ugly derivatives
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 15 36
Extensions
Hedged portfolios V
Download the Fama-French 3 factor + Momentum monthly data(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market compute (unhedged) Markowitzportfolio and Wald statistics
wstats lt- rbind(doboth(ff4xts[ 14]) wtrickws(ff4xts[
14] vcovfunc = sandwichvcovHAC))
rownames(wstats)[3] lt- c(weird trick w HAC)
xtable(wstats)
Mkt SMB HML UMD
Britten Jones t-stat 628 072 499 820weird trick Wald stat 537 077 447 603weird trick w HAC 510 077 392 553
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 16 36
Extensions
Hedged portfolios VI
Now hedge out Mkt
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
Gmat lt- matrix(diag(1 4)[1 ] ncol = 4)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 14] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD
Intercept 269 077 447 603
And compute the spanning Wald statistic
efstat lt- function(ism) ism$mu[1]sqrt(ism$Ohat[1 1])
print(efstat(asymv))
[1] 38
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 17 36
Extensions
Hedged portfolios VII
Now hedge out Mkt and RF
hedge out RFR too
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 15] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD RF
Intercept 071 205 227 343 -130
And the spanning statistic
print(efstat(asymv))
[1] 21
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 18 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
One Weird Trick
But is it useful
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 4 36
Portfolio Inference
Sample estimator
Since Θ = E[xxgt
]the simple estimator is unbiased
Θ =1
n
sum1leilen
xi xigt =
[1 microgt
micro Σ + micromicrogt
]
The inverse contains the sample estimates
Θminus1 =
1 + ζ2lowast minusνlowastgt
minusνlowast Σminus1
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 5 36
Portfolio Inference
Asymptotics I
By the Central Limit Theorem
radicn(
vech(
Θ)minus vech (Θ)
) N (0Ω)
where Ω = Var(vech
(xxgt
))
We can estimate Ω from the sample call it ΩItrsquos just sample covariance of vech
(xi xi
gt) for 1 le i le n
Use the delta method
radicn(
vech(
Θminus1)minus vech
(Θminus1
)) N
(0UΩUgt
)
Here U is some lsquouglyrsquo derivative depending on Θ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 6 36
Portfolio Inference
Asymptotics II
Ignoring details about symmetry etc the derivative is [7 12]
dXminus1
dX= minus
(Xminusgt otimes Xminus1
)
(This generalizes the scalar derivative)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 7 36
Portfolio Inference
I can make a hat or a brooch or a pterodactyl
Θminus1 =
1 + ζ2lowast minusνlowastgt
minusνlowast Σminus1
What is the use for Var
(vech
(Θminus1
))
Perform inference on elements of νlowast via Wald statistic(Compare elements of νlowast to their standard errors)
Perform inference on the maximal Sharpe ratio ζlowast
Equivalently Hotellingrsquos T 2 test (tests hypothesis micro is all zeros)
Portfolio shrinkage
Estimate the covariance of νlowast and Σminus1 (Attribute portfolio error toreturns or covariance) [5]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 8 36
Portfolio Inference
Implementation trust but verify
require(MarkowitzR)
setseed(2014)
X lt- matrix(rnorm(1000 5) ncol = 5) toy data
ism lt- MarkowitzRmp_vcov(X)
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
print(t(walds(ism))) Wald stats
X1 X2 X3 X4 X5
Intercept 089 -022 -16 -24 -049
cf Britten-Jones httpjstororgstable2697722
y lt- rep(1 dim(X)[1])
print(t(summary(lm(y ~ X - 1))$coefficients[ 3]))
X1 X2 X3 X4 X5
[1] 089 -022 -16 -25 -048
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 9 36
Portfolio Inference
Game over
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 10 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Extensions
Hedged portfolios I
Hedging the goal
Returns which are statistically independent from some random variables
Hedging a more realistic goal
A portfolio with zero covariance to some random variables
Hedging an achievable goal
A portfolio with zero sample covariance to some other portfolios oftradeable assets(eg you may have to hold some Mkt to hedge out the Mkt)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 12 36
Extensions
Hedged portfolios II
maxνGΣν=0νgtΣνleR2
νgtmicrominus r0radicνgtΣν
where G is a pg times p matrix of rank pg
Rows of G define portfolios against which we have 0 covariance
Typically G consists of some rows of identity matrix
ie ldquoMaximize Sharpe ratio with risk bound and zero covariance to someother portfoliosrdquoSolved by cνGlowast with c to satisfy risk bound and
νGlowast =(
Σminus1microminus Ggt(GΣGgt
)minus1Gmicro)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 13 36
Extensions
Hedged portfolios III
Use the weird trick Let G =
[1 00 G
] then
Θminus1 minus Ggt(
GΘGgt)minus1
G = microgtΣminus1microminus microgtGgt(GΣGgt
)minus1Gmicro minusνGlowast
gt
minusνGlowast Σminus1 minus Ggt(GΣGgt
)minus1G
minusνGlowast is the optimal hedged portfolio
UL corner is squared Sharpe ratio of νGlowastAlso used for portfolio spanning [16 8 10 11]
LR corner is loss of precision
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 14 36
Extensions
Hedged portfolios IV
Delta method gives the asymptotic distribution
radicn(
vech(
∆GΘminus1)minus vech
(∆GΘminus1
)) N
(0UΩUgt
)
with more ugly derivatives
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 15 36
Extensions
Hedged portfolios V
Download the Fama-French 3 factor + Momentum monthly data(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market compute (unhedged) Markowitzportfolio and Wald statistics
wstats lt- rbind(doboth(ff4xts[ 14]) wtrickws(ff4xts[
14] vcovfunc = sandwichvcovHAC))
rownames(wstats)[3] lt- c(weird trick w HAC)
xtable(wstats)
Mkt SMB HML UMD
Britten Jones t-stat 628 072 499 820weird trick Wald stat 537 077 447 603weird trick w HAC 510 077 392 553
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 16 36
Extensions
Hedged portfolios VI
Now hedge out Mkt
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
Gmat lt- matrix(diag(1 4)[1 ] ncol = 4)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 14] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD
Intercept 269 077 447 603
And compute the spanning Wald statistic
efstat lt- function(ism) ism$mu[1]sqrt(ism$Ohat[1 1])
print(efstat(asymv))
[1] 38
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 17 36
Extensions
Hedged portfolios VII
Now hedge out Mkt and RF
hedge out RFR too
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 15] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD RF
Intercept 071 205 227 343 -130
And the spanning statistic
print(efstat(asymv))
[1] 21
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 18 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Portfolio Inference
Sample estimator
Since Θ = E[xxgt
]the simple estimator is unbiased
Θ =1
n
sum1leilen
xi xigt =
[1 microgt
micro Σ + micromicrogt
]
The inverse contains the sample estimates
Θminus1 =
1 + ζ2lowast minusνlowastgt
minusνlowast Σminus1
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 5 36
Portfolio Inference
Asymptotics I
By the Central Limit Theorem
radicn(
vech(
Θ)minus vech (Θ)
) N (0Ω)
where Ω = Var(vech
(xxgt
))
We can estimate Ω from the sample call it ΩItrsquos just sample covariance of vech
(xi xi
gt) for 1 le i le n
Use the delta method
radicn(
vech(
Θminus1)minus vech
(Θminus1
)) N
(0UΩUgt
)
Here U is some lsquouglyrsquo derivative depending on Θ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 6 36
Portfolio Inference
Asymptotics II
Ignoring details about symmetry etc the derivative is [7 12]
dXminus1
dX= minus
(Xminusgt otimes Xminus1
)
(This generalizes the scalar derivative)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 7 36
Portfolio Inference
I can make a hat or a brooch or a pterodactyl
Θminus1 =
1 + ζ2lowast minusνlowastgt
minusνlowast Σminus1
What is the use for Var
(vech
(Θminus1
))
Perform inference on elements of νlowast via Wald statistic(Compare elements of νlowast to their standard errors)
Perform inference on the maximal Sharpe ratio ζlowast
Equivalently Hotellingrsquos T 2 test (tests hypothesis micro is all zeros)
Portfolio shrinkage
Estimate the covariance of νlowast and Σminus1 (Attribute portfolio error toreturns or covariance) [5]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 8 36
Portfolio Inference
Implementation trust but verify
require(MarkowitzR)
setseed(2014)
X lt- matrix(rnorm(1000 5) ncol = 5) toy data
ism lt- MarkowitzRmp_vcov(X)
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
print(t(walds(ism))) Wald stats
X1 X2 X3 X4 X5
Intercept 089 -022 -16 -24 -049
cf Britten-Jones httpjstororgstable2697722
y lt- rep(1 dim(X)[1])
print(t(summary(lm(y ~ X - 1))$coefficients[ 3]))
X1 X2 X3 X4 X5
[1] 089 -022 -16 -25 -048
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 9 36
Portfolio Inference
Game over
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 10 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Extensions
Hedged portfolios I
Hedging the goal
Returns which are statistically independent from some random variables
Hedging a more realistic goal
A portfolio with zero covariance to some random variables
Hedging an achievable goal
A portfolio with zero sample covariance to some other portfolios oftradeable assets(eg you may have to hold some Mkt to hedge out the Mkt)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 12 36
Extensions
Hedged portfolios II
maxνGΣν=0νgtΣνleR2
νgtmicrominus r0radicνgtΣν
where G is a pg times p matrix of rank pg
Rows of G define portfolios against which we have 0 covariance
Typically G consists of some rows of identity matrix
ie ldquoMaximize Sharpe ratio with risk bound and zero covariance to someother portfoliosrdquoSolved by cνGlowast with c to satisfy risk bound and
νGlowast =(
Σminus1microminus Ggt(GΣGgt
)minus1Gmicro)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 13 36
Extensions
Hedged portfolios III
Use the weird trick Let G =
[1 00 G
] then
Θminus1 minus Ggt(
GΘGgt)minus1
G = microgtΣminus1microminus microgtGgt(GΣGgt
)minus1Gmicro minusνGlowast
gt
minusνGlowast Σminus1 minus Ggt(GΣGgt
)minus1G
minusνGlowast is the optimal hedged portfolio
UL corner is squared Sharpe ratio of νGlowastAlso used for portfolio spanning [16 8 10 11]
LR corner is loss of precision
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 14 36
Extensions
Hedged portfolios IV
Delta method gives the asymptotic distribution
radicn(
vech(
∆GΘminus1)minus vech
(∆GΘminus1
)) N
(0UΩUgt
)
with more ugly derivatives
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 15 36
Extensions
Hedged portfolios V
Download the Fama-French 3 factor + Momentum monthly data(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market compute (unhedged) Markowitzportfolio and Wald statistics
wstats lt- rbind(doboth(ff4xts[ 14]) wtrickws(ff4xts[
14] vcovfunc = sandwichvcovHAC))
rownames(wstats)[3] lt- c(weird trick w HAC)
xtable(wstats)
Mkt SMB HML UMD
Britten Jones t-stat 628 072 499 820weird trick Wald stat 537 077 447 603weird trick w HAC 510 077 392 553
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 16 36
Extensions
Hedged portfolios VI
Now hedge out Mkt
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
Gmat lt- matrix(diag(1 4)[1 ] ncol = 4)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 14] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD
Intercept 269 077 447 603
And compute the spanning Wald statistic
efstat lt- function(ism) ism$mu[1]sqrt(ism$Ohat[1 1])
print(efstat(asymv))
[1] 38
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 17 36
Extensions
Hedged portfolios VII
Now hedge out Mkt and RF
hedge out RFR too
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 15] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD RF
Intercept 071 205 227 343 -130
And the spanning statistic
print(efstat(asymv))
[1] 21
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 18 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Portfolio Inference
Asymptotics I
By the Central Limit Theorem
radicn(
vech(
Θ)minus vech (Θ)
) N (0Ω)
where Ω = Var(vech
(xxgt
))
We can estimate Ω from the sample call it ΩItrsquos just sample covariance of vech
(xi xi
gt) for 1 le i le n
Use the delta method
radicn(
vech(
Θminus1)minus vech
(Θminus1
)) N
(0UΩUgt
)
Here U is some lsquouglyrsquo derivative depending on Θ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 6 36
Portfolio Inference
Asymptotics II
Ignoring details about symmetry etc the derivative is [7 12]
dXminus1
dX= minus
(Xminusgt otimes Xminus1
)
(This generalizes the scalar derivative)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 7 36
Portfolio Inference
I can make a hat or a brooch or a pterodactyl
Θminus1 =
1 + ζ2lowast minusνlowastgt
minusνlowast Σminus1
What is the use for Var
(vech
(Θminus1
))
Perform inference on elements of νlowast via Wald statistic(Compare elements of νlowast to their standard errors)
Perform inference on the maximal Sharpe ratio ζlowast
Equivalently Hotellingrsquos T 2 test (tests hypothesis micro is all zeros)
Portfolio shrinkage
Estimate the covariance of νlowast and Σminus1 (Attribute portfolio error toreturns or covariance) [5]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 8 36
Portfolio Inference
Implementation trust but verify
require(MarkowitzR)
setseed(2014)
X lt- matrix(rnorm(1000 5) ncol = 5) toy data
ism lt- MarkowitzRmp_vcov(X)
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
print(t(walds(ism))) Wald stats
X1 X2 X3 X4 X5
Intercept 089 -022 -16 -24 -049
cf Britten-Jones httpjstororgstable2697722
y lt- rep(1 dim(X)[1])
print(t(summary(lm(y ~ X - 1))$coefficients[ 3]))
X1 X2 X3 X4 X5
[1] 089 -022 -16 -25 -048
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 9 36
Portfolio Inference
Game over
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 10 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Extensions
Hedged portfolios I
Hedging the goal
Returns which are statistically independent from some random variables
Hedging a more realistic goal
A portfolio with zero covariance to some random variables
Hedging an achievable goal
A portfolio with zero sample covariance to some other portfolios oftradeable assets(eg you may have to hold some Mkt to hedge out the Mkt)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 12 36
Extensions
Hedged portfolios II
maxνGΣν=0νgtΣνleR2
νgtmicrominus r0radicνgtΣν
where G is a pg times p matrix of rank pg
Rows of G define portfolios against which we have 0 covariance
Typically G consists of some rows of identity matrix
ie ldquoMaximize Sharpe ratio with risk bound and zero covariance to someother portfoliosrdquoSolved by cνGlowast with c to satisfy risk bound and
νGlowast =(
Σminus1microminus Ggt(GΣGgt
)minus1Gmicro)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 13 36
Extensions
Hedged portfolios III
Use the weird trick Let G =
[1 00 G
] then
Θminus1 minus Ggt(
GΘGgt)minus1
G = microgtΣminus1microminus microgtGgt(GΣGgt
)minus1Gmicro minusνGlowast
gt
minusνGlowast Σminus1 minus Ggt(GΣGgt
)minus1G
minusνGlowast is the optimal hedged portfolio
UL corner is squared Sharpe ratio of νGlowastAlso used for portfolio spanning [16 8 10 11]
LR corner is loss of precision
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 14 36
Extensions
Hedged portfolios IV
Delta method gives the asymptotic distribution
radicn(
vech(
∆GΘminus1)minus vech
(∆GΘminus1
)) N
(0UΩUgt
)
with more ugly derivatives
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 15 36
Extensions
Hedged portfolios V
Download the Fama-French 3 factor + Momentum monthly data(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market compute (unhedged) Markowitzportfolio and Wald statistics
wstats lt- rbind(doboth(ff4xts[ 14]) wtrickws(ff4xts[
14] vcovfunc = sandwichvcovHAC))
rownames(wstats)[3] lt- c(weird trick w HAC)
xtable(wstats)
Mkt SMB HML UMD
Britten Jones t-stat 628 072 499 820weird trick Wald stat 537 077 447 603weird trick w HAC 510 077 392 553
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 16 36
Extensions
Hedged portfolios VI
Now hedge out Mkt
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
Gmat lt- matrix(diag(1 4)[1 ] ncol = 4)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 14] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD
Intercept 269 077 447 603
And compute the spanning Wald statistic
efstat lt- function(ism) ism$mu[1]sqrt(ism$Ohat[1 1])
print(efstat(asymv))
[1] 38
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 17 36
Extensions
Hedged portfolios VII
Now hedge out Mkt and RF
hedge out RFR too
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 15] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD RF
Intercept 071 205 227 343 -130
And the spanning statistic
print(efstat(asymv))
[1] 21
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 18 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Portfolio Inference
Asymptotics II
Ignoring details about symmetry etc the derivative is [7 12]
dXminus1
dX= minus
(Xminusgt otimes Xminus1
)
(This generalizes the scalar derivative)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 7 36
Portfolio Inference
I can make a hat or a brooch or a pterodactyl
Θminus1 =
1 + ζ2lowast minusνlowastgt
minusνlowast Σminus1
What is the use for Var
(vech
(Θminus1
))
Perform inference on elements of νlowast via Wald statistic(Compare elements of νlowast to their standard errors)
Perform inference on the maximal Sharpe ratio ζlowast
Equivalently Hotellingrsquos T 2 test (tests hypothesis micro is all zeros)
Portfolio shrinkage
Estimate the covariance of νlowast and Σminus1 (Attribute portfolio error toreturns or covariance) [5]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 8 36
Portfolio Inference
Implementation trust but verify
require(MarkowitzR)
setseed(2014)
X lt- matrix(rnorm(1000 5) ncol = 5) toy data
ism lt- MarkowitzRmp_vcov(X)
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
print(t(walds(ism))) Wald stats
X1 X2 X3 X4 X5
Intercept 089 -022 -16 -24 -049
cf Britten-Jones httpjstororgstable2697722
y lt- rep(1 dim(X)[1])
print(t(summary(lm(y ~ X - 1))$coefficients[ 3]))
X1 X2 X3 X4 X5
[1] 089 -022 -16 -25 -048
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 9 36
Portfolio Inference
Game over
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 10 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Extensions
Hedged portfolios I
Hedging the goal
Returns which are statistically independent from some random variables
Hedging a more realistic goal
A portfolio with zero covariance to some random variables
Hedging an achievable goal
A portfolio with zero sample covariance to some other portfolios oftradeable assets(eg you may have to hold some Mkt to hedge out the Mkt)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 12 36
Extensions
Hedged portfolios II
maxνGΣν=0νgtΣνleR2
νgtmicrominus r0radicνgtΣν
where G is a pg times p matrix of rank pg
Rows of G define portfolios against which we have 0 covariance
Typically G consists of some rows of identity matrix
ie ldquoMaximize Sharpe ratio with risk bound and zero covariance to someother portfoliosrdquoSolved by cνGlowast with c to satisfy risk bound and
νGlowast =(
Σminus1microminus Ggt(GΣGgt
)minus1Gmicro)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 13 36
Extensions
Hedged portfolios III
Use the weird trick Let G =
[1 00 G
] then
Θminus1 minus Ggt(
GΘGgt)minus1
G = microgtΣminus1microminus microgtGgt(GΣGgt
)minus1Gmicro minusνGlowast
gt
minusνGlowast Σminus1 minus Ggt(GΣGgt
)minus1G
minusνGlowast is the optimal hedged portfolio
UL corner is squared Sharpe ratio of νGlowastAlso used for portfolio spanning [16 8 10 11]
LR corner is loss of precision
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 14 36
Extensions
Hedged portfolios IV
Delta method gives the asymptotic distribution
radicn(
vech(
∆GΘminus1)minus vech
(∆GΘminus1
)) N
(0UΩUgt
)
with more ugly derivatives
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 15 36
Extensions
Hedged portfolios V
Download the Fama-French 3 factor + Momentum monthly data(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market compute (unhedged) Markowitzportfolio and Wald statistics
wstats lt- rbind(doboth(ff4xts[ 14]) wtrickws(ff4xts[
14] vcovfunc = sandwichvcovHAC))
rownames(wstats)[3] lt- c(weird trick w HAC)
xtable(wstats)
Mkt SMB HML UMD
Britten Jones t-stat 628 072 499 820weird trick Wald stat 537 077 447 603weird trick w HAC 510 077 392 553
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 16 36
Extensions
Hedged portfolios VI
Now hedge out Mkt
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
Gmat lt- matrix(diag(1 4)[1 ] ncol = 4)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 14] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD
Intercept 269 077 447 603
And compute the spanning Wald statistic
efstat lt- function(ism) ism$mu[1]sqrt(ism$Ohat[1 1])
print(efstat(asymv))
[1] 38
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 17 36
Extensions
Hedged portfolios VII
Now hedge out Mkt and RF
hedge out RFR too
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 15] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD RF
Intercept 071 205 227 343 -130
And the spanning statistic
print(efstat(asymv))
[1] 21
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 18 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Portfolio Inference
I can make a hat or a brooch or a pterodactyl
Θminus1 =
1 + ζ2lowast minusνlowastgt
minusνlowast Σminus1
What is the use for Var
(vech
(Θminus1
))
Perform inference on elements of νlowast via Wald statistic(Compare elements of νlowast to their standard errors)
Perform inference on the maximal Sharpe ratio ζlowast
Equivalently Hotellingrsquos T 2 test (tests hypothesis micro is all zeros)
Portfolio shrinkage
Estimate the covariance of νlowast and Σminus1 (Attribute portfolio error toreturns or covariance) [5]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 8 36
Portfolio Inference
Implementation trust but verify
require(MarkowitzR)
setseed(2014)
X lt- matrix(rnorm(1000 5) ncol = 5) toy data
ism lt- MarkowitzRmp_vcov(X)
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
print(t(walds(ism))) Wald stats
X1 X2 X3 X4 X5
Intercept 089 -022 -16 -24 -049
cf Britten-Jones httpjstororgstable2697722
y lt- rep(1 dim(X)[1])
print(t(summary(lm(y ~ X - 1))$coefficients[ 3]))
X1 X2 X3 X4 X5
[1] 089 -022 -16 -25 -048
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 9 36
Portfolio Inference
Game over
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 10 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Extensions
Hedged portfolios I
Hedging the goal
Returns which are statistically independent from some random variables
Hedging a more realistic goal
A portfolio with zero covariance to some random variables
Hedging an achievable goal
A portfolio with zero sample covariance to some other portfolios oftradeable assets(eg you may have to hold some Mkt to hedge out the Mkt)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 12 36
Extensions
Hedged portfolios II
maxνGΣν=0νgtΣνleR2
νgtmicrominus r0radicνgtΣν
where G is a pg times p matrix of rank pg
Rows of G define portfolios against which we have 0 covariance
Typically G consists of some rows of identity matrix
ie ldquoMaximize Sharpe ratio with risk bound and zero covariance to someother portfoliosrdquoSolved by cνGlowast with c to satisfy risk bound and
νGlowast =(
Σminus1microminus Ggt(GΣGgt
)minus1Gmicro)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 13 36
Extensions
Hedged portfolios III
Use the weird trick Let G =
[1 00 G
] then
Θminus1 minus Ggt(
GΘGgt)minus1
G = microgtΣminus1microminus microgtGgt(GΣGgt
)minus1Gmicro minusνGlowast
gt
minusνGlowast Σminus1 minus Ggt(GΣGgt
)minus1G
minusνGlowast is the optimal hedged portfolio
UL corner is squared Sharpe ratio of νGlowastAlso used for portfolio spanning [16 8 10 11]
LR corner is loss of precision
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 14 36
Extensions
Hedged portfolios IV
Delta method gives the asymptotic distribution
radicn(
vech(
∆GΘminus1)minus vech
(∆GΘminus1
)) N
(0UΩUgt
)
with more ugly derivatives
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 15 36
Extensions
Hedged portfolios V
Download the Fama-French 3 factor + Momentum monthly data(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market compute (unhedged) Markowitzportfolio and Wald statistics
wstats lt- rbind(doboth(ff4xts[ 14]) wtrickws(ff4xts[
14] vcovfunc = sandwichvcovHAC))
rownames(wstats)[3] lt- c(weird trick w HAC)
xtable(wstats)
Mkt SMB HML UMD
Britten Jones t-stat 628 072 499 820weird trick Wald stat 537 077 447 603weird trick w HAC 510 077 392 553
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 16 36
Extensions
Hedged portfolios VI
Now hedge out Mkt
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
Gmat lt- matrix(diag(1 4)[1 ] ncol = 4)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 14] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD
Intercept 269 077 447 603
And compute the spanning Wald statistic
efstat lt- function(ism) ism$mu[1]sqrt(ism$Ohat[1 1])
print(efstat(asymv))
[1] 38
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 17 36
Extensions
Hedged portfolios VII
Now hedge out Mkt and RF
hedge out RFR too
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 15] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD RF
Intercept 071 205 227 343 -130
And the spanning statistic
print(efstat(asymv))
[1] 21
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 18 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Portfolio Inference
Implementation trust but verify
require(MarkowitzR)
setseed(2014)
X lt- matrix(rnorm(1000 5) ncol = 5) toy data
ism lt- MarkowitzRmp_vcov(X)
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
print(t(walds(ism))) Wald stats
X1 X2 X3 X4 X5
Intercept 089 -022 -16 -24 -049
cf Britten-Jones httpjstororgstable2697722
y lt- rep(1 dim(X)[1])
print(t(summary(lm(y ~ X - 1))$coefficients[ 3]))
X1 X2 X3 X4 X5
[1] 089 -022 -16 -25 -048
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 9 36
Portfolio Inference
Game over
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 10 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Extensions
Hedged portfolios I
Hedging the goal
Returns which are statistically independent from some random variables
Hedging a more realistic goal
A portfolio with zero covariance to some random variables
Hedging an achievable goal
A portfolio with zero sample covariance to some other portfolios oftradeable assets(eg you may have to hold some Mkt to hedge out the Mkt)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 12 36
Extensions
Hedged portfolios II
maxνGΣν=0νgtΣνleR2
νgtmicrominus r0radicνgtΣν
where G is a pg times p matrix of rank pg
Rows of G define portfolios against which we have 0 covariance
Typically G consists of some rows of identity matrix
ie ldquoMaximize Sharpe ratio with risk bound and zero covariance to someother portfoliosrdquoSolved by cνGlowast with c to satisfy risk bound and
νGlowast =(
Σminus1microminus Ggt(GΣGgt
)minus1Gmicro)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 13 36
Extensions
Hedged portfolios III
Use the weird trick Let G =
[1 00 G
] then
Θminus1 minus Ggt(
GΘGgt)minus1
G = microgtΣminus1microminus microgtGgt(GΣGgt
)minus1Gmicro minusνGlowast
gt
minusνGlowast Σminus1 minus Ggt(GΣGgt
)minus1G
minusνGlowast is the optimal hedged portfolio
UL corner is squared Sharpe ratio of νGlowastAlso used for portfolio spanning [16 8 10 11]
LR corner is loss of precision
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 14 36
Extensions
Hedged portfolios IV
Delta method gives the asymptotic distribution
radicn(
vech(
∆GΘminus1)minus vech
(∆GΘminus1
)) N
(0UΩUgt
)
with more ugly derivatives
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 15 36
Extensions
Hedged portfolios V
Download the Fama-French 3 factor + Momentum monthly data(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market compute (unhedged) Markowitzportfolio and Wald statistics
wstats lt- rbind(doboth(ff4xts[ 14]) wtrickws(ff4xts[
14] vcovfunc = sandwichvcovHAC))
rownames(wstats)[3] lt- c(weird trick w HAC)
xtable(wstats)
Mkt SMB HML UMD
Britten Jones t-stat 628 072 499 820weird trick Wald stat 537 077 447 603weird trick w HAC 510 077 392 553
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 16 36
Extensions
Hedged portfolios VI
Now hedge out Mkt
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
Gmat lt- matrix(diag(1 4)[1 ] ncol = 4)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 14] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD
Intercept 269 077 447 603
And compute the spanning Wald statistic
efstat lt- function(ism) ism$mu[1]sqrt(ism$Ohat[1 1])
print(efstat(asymv))
[1] 38
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 17 36
Extensions
Hedged portfolios VII
Now hedge out Mkt and RF
hedge out RFR too
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 15] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD RF
Intercept 071 205 227 343 -130
And the spanning statistic
print(efstat(asymv))
[1] 21
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 18 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Portfolio Inference
Game over
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 10 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Extensions
Hedged portfolios I
Hedging the goal
Returns which are statistically independent from some random variables
Hedging a more realistic goal
A portfolio with zero covariance to some random variables
Hedging an achievable goal
A portfolio with zero sample covariance to some other portfolios oftradeable assets(eg you may have to hold some Mkt to hedge out the Mkt)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 12 36
Extensions
Hedged portfolios II
maxνGΣν=0νgtΣνleR2
νgtmicrominus r0radicνgtΣν
where G is a pg times p matrix of rank pg
Rows of G define portfolios against which we have 0 covariance
Typically G consists of some rows of identity matrix
ie ldquoMaximize Sharpe ratio with risk bound and zero covariance to someother portfoliosrdquoSolved by cνGlowast with c to satisfy risk bound and
νGlowast =(
Σminus1microminus Ggt(GΣGgt
)minus1Gmicro)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 13 36
Extensions
Hedged portfolios III
Use the weird trick Let G =
[1 00 G
] then
Θminus1 minus Ggt(
GΘGgt)minus1
G = microgtΣminus1microminus microgtGgt(GΣGgt
)minus1Gmicro minusνGlowast
gt
minusνGlowast Σminus1 minus Ggt(GΣGgt
)minus1G
minusνGlowast is the optimal hedged portfolio
UL corner is squared Sharpe ratio of νGlowastAlso used for portfolio spanning [16 8 10 11]
LR corner is loss of precision
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 14 36
Extensions
Hedged portfolios IV
Delta method gives the asymptotic distribution
radicn(
vech(
∆GΘminus1)minus vech
(∆GΘminus1
)) N
(0UΩUgt
)
with more ugly derivatives
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 15 36
Extensions
Hedged portfolios V
Download the Fama-French 3 factor + Momentum monthly data(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market compute (unhedged) Markowitzportfolio and Wald statistics
wstats lt- rbind(doboth(ff4xts[ 14]) wtrickws(ff4xts[
14] vcovfunc = sandwichvcovHAC))
rownames(wstats)[3] lt- c(weird trick w HAC)
xtable(wstats)
Mkt SMB HML UMD
Britten Jones t-stat 628 072 499 820weird trick Wald stat 537 077 447 603weird trick w HAC 510 077 392 553
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 16 36
Extensions
Hedged portfolios VI
Now hedge out Mkt
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
Gmat lt- matrix(diag(1 4)[1 ] ncol = 4)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 14] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD
Intercept 269 077 447 603
And compute the spanning Wald statistic
efstat lt- function(ism) ism$mu[1]sqrt(ism$Ohat[1 1])
print(efstat(asymv))
[1] 38
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 17 36
Extensions
Hedged portfolios VII
Now hedge out Mkt and RF
hedge out RFR too
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 15] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD RF
Intercept 071 205 227 343 -130
And the spanning statistic
print(efstat(asymv))
[1] 21
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 18 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Extensions
Hedged portfolios I
Hedging the goal
Returns which are statistically independent from some random variables
Hedging a more realistic goal
A portfolio with zero covariance to some random variables
Hedging an achievable goal
A portfolio with zero sample covariance to some other portfolios oftradeable assets(eg you may have to hold some Mkt to hedge out the Mkt)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 12 36
Extensions
Hedged portfolios II
maxνGΣν=0νgtΣνleR2
νgtmicrominus r0radicνgtΣν
where G is a pg times p matrix of rank pg
Rows of G define portfolios against which we have 0 covariance
Typically G consists of some rows of identity matrix
ie ldquoMaximize Sharpe ratio with risk bound and zero covariance to someother portfoliosrdquoSolved by cνGlowast with c to satisfy risk bound and
νGlowast =(
Σminus1microminus Ggt(GΣGgt
)minus1Gmicro)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 13 36
Extensions
Hedged portfolios III
Use the weird trick Let G =
[1 00 G
] then
Θminus1 minus Ggt(
GΘGgt)minus1
G = microgtΣminus1microminus microgtGgt(GΣGgt
)minus1Gmicro minusνGlowast
gt
minusνGlowast Σminus1 minus Ggt(GΣGgt
)minus1G
minusνGlowast is the optimal hedged portfolio
UL corner is squared Sharpe ratio of νGlowastAlso used for portfolio spanning [16 8 10 11]
LR corner is loss of precision
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 14 36
Extensions
Hedged portfolios IV
Delta method gives the asymptotic distribution
radicn(
vech(
∆GΘminus1)minus vech
(∆GΘminus1
)) N
(0UΩUgt
)
with more ugly derivatives
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 15 36
Extensions
Hedged portfolios V
Download the Fama-French 3 factor + Momentum monthly data(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market compute (unhedged) Markowitzportfolio and Wald statistics
wstats lt- rbind(doboth(ff4xts[ 14]) wtrickws(ff4xts[
14] vcovfunc = sandwichvcovHAC))
rownames(wstats)[3] lt- c(weird trick w HAC)
xtable(wstats)
Mkt SMB HML UMD
Britten Jones t-stat 628 072 499 820weird trick Wald stat 537 077 447 603weird trick w HAC 510 077 392 553
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 16 36
Extensions
Hedged portfolios VI
Now hedge out Mkt
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
Gmat lt- matrix(diag(1 4)[1 ] ncol = 4)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 14] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD
Intercept 269 077 447 603
And compute the spanning Wald statistic
efstat lt- function(ism) ism$mu[1]sqrt(ism$Ohat[1 1])
print(efstat(asymv))
[1] 38
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 17 36
Extensions
Hedged portfolios VII
Now hedge out Mkt and RF
hedge out RFR too
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 15] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD RF
Intercept 071 205 227 343 -130
And the spanning statistic
print(efstat(asymv))
[1] 21
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 18 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Portfolio Inference
Selling this weird trick
Why weird trick not Britten-Jones or Okhrin et al [4 2 14]
Fewer assumptions fourth moments exist vs normality of returns
Straightforward to use HAC estimator for Ω
Models covariance between return and volatility (At a cost)
Solves a larger problem eg can use for inference on ζ2lowast
Real question whatrsquos wrong with vanilla Markowitz
This trick can be adapted to deal with
Hedged portfolios
Heteroskedasticity
Conditional expected returns
Perhaps more
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 11 36
Extensions
Hedged portfolios I
Hedging the goal
Returns which are statistically independent from some random variables
Hedging a more realistic goal
A portfolio with zero covariance to some random variables
Hedging an achievable goal
A portfolio with zero sample covariance to some other portfolios oftradeable assets(eg you may have to hold some Mkt to hedge out the Mkt)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 12 36
Extensions
Hedged portfolios II
maxνGΣν=0νgtΣνleR2
νgtmicrominus r0radicνgtΣν
where G is a pg times p matrix of rank pg
Rows of G define portfolios against which we have 0 covariance
Typically G consists of some rows of identity matrix
ie ldquoMaximize Sharpe ratio with risk bound and zero covariance to someother portfoliosrdquoSolved by cνGlowast with c to satisfy risk bound and
νGlowast =(
Σminus1microminus Ggt(GΣGgt
)minus1Gmicro)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 13 36
Extensions
Hedged portfolios III
Use the weird trick Let G =
[1 00 G
] then
Θminus1 minus Ggt(
GΘGgt)minus1
G = microgtΣminus1microminus microgtGgt(GΣGgt
)minus1Gmicro minusνGlowast
gt
minusνGlowast Σminus1 minus Ggt(GΣGgt
)minus1G
minusνGlowast is the optimal hedged portfolio
UL corner is squared Sharpe ratio of νGlowastAlso used for portfolio spanning [16 8 10 11]
LR corner is loss of precision
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 14 36
Extensions
Hedged portfolios IV
Delta method gives the asymptotic distribution
radicn(
vech(
∆GΘminus1)minus vech
(∆GΘminus1
)) N
(0UΩUgt
)
with more ugly derivatives
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 15 36
Extensions
Hedged portfolios V
Download the Fama-French 3 factor + Momentum monthly data(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market compute (unhedged) Markowitzportfolio and Wald statistics
wstats lt- rbind(doboth(ff4xts[ 14]) wtrickws(ff4xts[
14] vcovfunc = sandwichvcovHAC))
rownames(wstats)[3] lt- c(weird trick w HAC)
xtable(wstats)
Mkt SMB HML UMD
Britten Jones t-stat 628 072 499 820weird trick Wald stat 537 077 447 603weird trick w HAC 510 077 392 553
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 16 36
Extensions
Hedged portfolios VI
Now hedge out Mkt
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
Gmat lt- matrix(diag(1 4)[1 ] ncol = 4)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 14] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD
Intercept 269 077 447 603
And compute the spanning Wald statistic
efstat lt- function(ism) ism$mu[1]sqrt(ism$Ohat[1 1])
print(efstat(asymv))
[1] 38
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 17 36
Extensions
Hedged portfolios VII
Now hedge out Mkt and RF
hedge out RFR too
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 15] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD RF
Intercept 071 205 227 343 -130
And the spanning statistic
print(efstat(asymv))
[1] 21
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 18 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Extensions
Hedged portfolios I
Hedging the goal
Returns which are statistically independent from some random variables
Hedging a more realistic goal
A portfolio with zero covariance to some random variables
Hedging an achievable goal
A portfolio with zero sample covariance to some other portfolios oftradeable assets(eg you may have to hold some Mkt to hedge out the Mkt)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 12 36
Extensions
Hedged portfolios II
maxνGΣν=0νgtΣνleR2
νgtmicrominus r0radicνgtΣν
where G is a pg times p matrix of rank pg
Rows of G define portfolios against which we have 0 covariance
Typically G consists of some rows of identity matrix
ie ldquoMaximize Sharpe ratio with risk bound and zero covariance to someother portfoliosrdquoSolved by cνGlowast with c to satisfy risk bound and
νGlowast =(
Σminus1microminus Ggt(GΣGgt
)minus1Gmicro)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 13 36
Extensions
Hedged portfolios III
Use the weird trick Let G =
[1 00 G
] then
Θminus1 minus Ggt(
GΘGgt)minus1
G = microgtΣminus1microminus microgtGgt(GΣGgt
)minus1Gmicro minusνGlowast
gt
minusνGlowast Σminus1 minus Ggt(GΣGgt
)minus1G
minusνGlowast is the optimal hedged portfolio
UL corner is squared Sharpe ratio of νGlowastAlso used for portfolio spanning [16 8 10 11]
LR corner is loss of precision
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 14 36
Extensions
Hedged portfolios IV
Delta method gives the asymptotic distribution
radicn(
vech(
∆GΘminus1)minus vech
(∆GΘminus1
)) N
(0UΩUgt
)
with more ugly derivatives
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 15 36
Extensions
Hedged portfolios V
Download the Fama-French 3 factor + Momentum monthly data(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market compute (unhedged) Markowitzportfolio and Wald statistics
wstats lt- rbind(doboth(ff4xts[ 14]) wtrickws(ff4xts[
14] vcovfunc = sandwichvcovHAC))
rownames(wstats)[3] lt- c(weird trick w HAC)
xtable(wstats)
Mkt SMB HML UMD
Britten Jones t-stat 628 072 499 820weird trick Wald stat 537 077 447 603weird trick w HAC 510 077 392 553
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 16 36
Extensions
Hedged portfolios VI
Now hedge out Mkt
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
Gmat lt- matrix(diag(1 4)[1 ] ncol = 4)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 14] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD
Intercept 269 077 447 603
And compute the spanning Wald statistic
efstat lt- function(ism) ism$mu[1]sqrt(ism$Ohat[1 1])
print(efstat(asymv))
[1] 38
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 17 36
Extensions
Hedged portfolios VII
Now hedge out Mkt and RF
hedge out RFR too
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 15] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD RF
Intercept 071 205 227 343 -130
And the spanning statistic
print(efstat(asymv))
[1] 21
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 18 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Extensions
Hedged portfolios II
maxνGΣν=0νgtΣνleR2
νgtmicrominus r0radicνgtΣν
where G is a pg times p matrix of rank pg
Rows of G define portfolios against which we have 0 covariance
Typically G consists of some rows of identity matrix
ie ldquoMaximize Sharpe ratio with risk bound and zero covariance to someother portfoliosrdquoSolved by cνGlowast with c to satisfy risk bound and
νGlowast =(
Σminus1microminus Ggt(GΣGgt
)minus1Gmicro)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 13 36
Extensions
Hedged portfolios III
Use the weird trick Let G =
[1 00 G
] then
Θminus1 minus Ggt(
GΘGgt)minus1
G = microgtΣminus1microminus microgtGgt(GΣGgt
)minus1Gmicro minusνGlowast
gt
minusνGlowast Σminus1 minus Ggt(GΣGgt
)minus1G
minusνGlowast is the optimal hedged portfolio
UL corner is squared Sharpe ratio of νGlowastAlso used for portfolio spanning [16 8 10 11]
LR corner is loss of precision
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 14 36
Extensions
Hedged portfolios IV
Delta method gives the asymptotic distribution
radicn(
vech(
∆GΘminus1)minus vech
(∆GΘminus1
)) N
(0UΩUgt
)
with more ugly derivatives
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 15 36
Extensions
Hedged portfolios V
Download the Fama-French 3 factor + Momentum monthly data(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market compute (unhedged) Markowitzportfolio and Wald statistics
wstats lt- rbind(doboth(ff4xts[ 14]) wtrickws(ff4xts[
14] vcovfunc = sandwichvcovHAC))
rownames(wstats)[3] lt- c(weird trick w HAC)
xtable(wstats)
Mkt SMB HML UMD
Britten Jones t-stat 628 072 499 820weird trick Wald stat 537 077 447 603weird trick w HAC 510 077 392 553
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 16 36
Extensions
Hedged portfolios VI
Now hedge out Mkt
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
Gmat lt- matrix(diag(1 4)[1 ] ncol = 4)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 14] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD
Intercept 269 077 447 603
And compute the spanning Wald statistic
efstat lt- function(ism) ism$mu[1]sqrt(ism$Ohat[1 1])
print(efstat(asymv))
[1] 38
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 17 36
Extensions
Hedged portfolios VII
Now hedge out Mkt and RF
hedge out RFR too
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 15] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD RF
Intercept 071 205 227 343 -130
And the spanning statistic
print(efstat(asymv))
[1] 21
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 18 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Extensions
Hedged portfolios III
Use the weird trick Let G =
[1 00 G
] then
Θminus1 minus Ggt(
GΘGgt)minus1
G = microgtΣminus1microminus microgtGgt(GΣGgt
)minus1Gmicro minusνGlowast
gt
minusνGlowast Σminus1 minus Ggt(GΣGgt
)minus1G
minusνGlowast is the optimal hedged portfolio
UL corner is squared Sharpe ratio of νGlowastAlso used for portfolio spanning [16 8 10 11]
LR corner is loss of precision
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 14 36
Extensions
Hedged portfolios IV
Delta method gives the asymptotic distribution
radicn(
vech(
∆GΘminus1)minus vech
(∆GΘminus1
)) N
(0UΩUgt
)
with more ugly derivatives
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 15 36
Extensions
Hedged portfolios V
Download the Fama-French 3 factor + Momentum monthly data(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market compute (unhedged) Markowitzportfolio and Wald statistics
wstats lt- rbind(doboth(ff4xts[ 14]) wtrickws(ff4xts[
14] vcovfunc = sandwichvcovHAC))
rownames(wstats)[3] lt- c(weird trick w HAC)
xtable(wstats)
Mkt SMB HML UMD
Britten Jones t-stat 628 072 499 820weird trick Wald stat 537 077 447 603weird trick w HAC 510 077 392 553
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 16 36
Extensions
Hedged portfolios VI
Now hedge out Mkt
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
Gmat lt- matrix(diag(1 4)[1 ] ncol = 4)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 14] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD
Intercept 269 077 447 603
And compute the spanning Wald statistic
efstat lt- function(ism) ism$mu[1]sqrt(ism$Ohat[1 1])
print(efstat(asymv))
[1] 38
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 17 36
Extensions
Hedged portfolios VII
Now hedge out Mkt and RF
hedge out RFR too
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 15] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD RF
Intercept 071 205 227 343 -130
And the spanning statistic
print(efstat(asymv))
[1] 21
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 18 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Extensions
Hedged portfolios IV
Delta method gives the asymptotic distribution
radicn(
vech(
∆GΘminus1)minus vech
(∆GΘminus1
)) N
(0UΩUgt
)
with more ugly derivatives
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 15 36
Extensions
Hedged portfolios V
Download the Fama-French 3 factor + Momentum monthly data(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market compute (unhedged) Markowitzportfolio and Wald statistics
wstats lt- rbind(doboth(ff4xts[ 14]) wtrickws(ff4xts[
14] vcovfunc = sandwichvcovHAC))
rownames(wstats)[3] lt- c(weird trick w HAC)
xtable(wstats)
Mkt SMB HML UMD
Britten Jones t-stat 628 072 499 820weird trick Wald stat 537 077 447 603weird trick w HAC 510 077 392 553
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 16 36
Extensions
Hedged portfolios VI
Now hedge out Mkt
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
Gmat lt- matrix(diag(1 4)[1 ] ncol = 4)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 14] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD
Intercept 269 077 447 603
And compute the spanning Wald statistic
efstat lt- function(ism) ism$mu[1]sqrt(ism$Ohat[1 1])
print(efstat(asymv))
[1] 38
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 17 36
Extensions
Hedged portfolios VII
Now hedge out Mkt and RF
hedge out RFR too
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 15] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD RF
Intercept 071 205 227 343 -130
And the spanning statistic
print(efstat(asymv))
[1] 21
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 18 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Extensions
Hedged portfolios V
Download the Fama-French 3 factor + Momentum monthly data(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market compute (unhedged) Markowitzportfolio and Wald statistics
wstats lt- rbind(doboth(ff4xts[ 14]) wtrickws(ff4xts[
14] vcovfunc = sandwichvcovHAC))
rownames(wstats)[3] lt- c(weird trick w HAC)
xtable(wstats)
Mkt SMB HML UMD
Britten Jones t-stat 628 072 499 820weird trick Wald stat 537 077 447 603weird trick w HAC 510 077 392 553
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 16 36
Extensions
Hedged portfolios VI
Now hedge out Mkt
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
Gmat lt- matrix(diag(1 4)[1 ] ncol = 4)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 14] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD
Intercept 269 077 447 603
And compute the spanning Wald statistic
efstat lt- function(ism) ism$mu[1]sqrt(ism$Ohat[1 1])
print(efstat(asymv))
[1] 38
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 17 36
Extensions
Hedged portfolios VII
Now hedge out Mkt and RF
hedge out RFR too
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 15] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD RF
Intercept 071 205 227 343 -130
And the spanning statistic
print(efstat(asymv))
[1] 21
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 18 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Extensions
Hedged portfolios VI
Now hedge out Mkt
walds lt- function(ism) ism$Wsqrt(diag(ism$What))
Gmat lt- matrix(diag(1 4)[1 ] ncol = 4)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 14] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD
Intercept 269 077 447 603
And compute the spanning Wald statistic
efstat lt- function(ism) ism$mu[1]sqrt(ism$Ohat[1 1])
print(efstat(asymv))
[1] 38
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 17 36
Extensions
Hedged portfolios VII
Now hedge out Mkt and RF
hedge out RFR too
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 15] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD RF
Intercept 071 205 227 343 -130
And the spanning statistic
print(efstat(asymv))
[1] 21
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 18 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Extensions
Hedged portfolios VII
Now hedge out Mkt and RF
hedge out RFR too
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asymv lt- MarkowitzRmp_vcov(ff4xts[ 15] fitintercept = TRUE
Gmat = Gmat)
xtable(t(walds(asymv)))
Mkt SMB HML UMD RF
Intercept 071 205 227 343 -130
And the spanning statistic
print(efstat(asymv))
[1] 21
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 18 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Extensions
Heteroskedasticity
Prior to investment decision observe si proportional to volatility
Two competing models
(constant) E [xi+1 | si ] = simicro Var (xi+1 | si ) = si2Σ
(floating) E [xi+1 | si ] = micro Var (xi+1 | si ) = si2Σ
For (constant) ζlowast isradicmicrogtΣminus1micro independent of si (Volatility time
vs wall-clock time)For (floating) it is si
minus1radicmicrogtΣminus1micro higher when volatility is low
(Volatility drinks your milkshake)
Why do I have to choose
(mixed) E [xi+1 | si ] = simicro0 + micro1 Var (xi+1 | si ) = si2Σ
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 19 36
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Extensions
Conditional expectation I
Suppose f -vector f i observed prior to investment decision and
(conditional) E [xi+1 | f i ] = Bf i Var (xi+1 | f i ) = Σ
B is some p times f matrix [6 9 3]
Conditional on observing f i solve
argmaxν Var(νgtxi+1| f i )leR2
E[νgtxi+1 | f i
]minus r0radic
Var (νgtxi+1 | f i )
for r0 ge 0R gt 0ldquoMaximize Sharpe with bound on risk conditional on f i rdquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 20 36
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Extensions
Conditional expectation II
Optimal portfolio is cνlowast with
νlowast = Σminus1B f i
Σminus1B generalizes the Markowitz portfoliothe coefficient of the Sharpe-optimal portfolio linear in features f i The lsquoMarkowitz coefficientrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 21 36
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Extensions
Conditional expectation III
Same weird trick works Let ˜xi+1 =[f igt xi+1
gt]gtThe uncentered second moment is
Θf = E[˜x˜xgt]
=
[Γf Γf Bgt
BΓf Σ + BΓf Bgt
] where Γf = E
[ffgt]
The inverse of Θf is
Θfminus1 =
[Γfminus1 + BgtΣminus1B minusBgtΣminus1
minusΣminus1B Σminus1
]
Σminus1B appears in off diagonals
BgtΣminus1B related to HLT
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 22 36
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Extensions
Conditional expectation IV
Again define sample estimator
Θf =1
n
sum1leilen
˜xi˜xigt
Use Central Limit theorem and delta method to get
radicn(
vech(
Θminus1f
)minus vech
(Θfminus1)) N
(0UΩUgt
)
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 23 36
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Examples
Examples I
Take the Fama-French 3 factor + Momentum monthly returns(1927-02-01 to 2014-01-01) from Quandl [13]
Add risk-free rate back to market
Use Shillerrsquos PE ratio as predictive state variable
Z-score the PE data
zsc lt- function(x ) (x - mean(x ))sd(x )
featuresz lt- zsc(features narm = TRUE)
asym lt- MarkowitzRmp_vcov(ff4xts[ 14] featuresz
fitintercept = TRUE vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD
Intercept 310 330 240 350Cyclically Adjusted PE Ratio -180 -100 -009 370
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 24 36
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Examples
Examples II
Now the same but hedge out Mkt and RF
hedge out Mkt and RF
Gmat lt- matrix(diag(1 5)[c(1 5) ] ncol = 5)
asym lt- MarkowitzRmp_vcov(ff4xts[ 15] featuresz
fitintercept = TRUE Gmat = Gmat vcovfunc = sandwichvcovHAC)
xtable(signif(t(walds(asym)) digits = 2))
Mkt SMB HML UMD RF
Intercept 055 210 210 210 -150Cyclically Adjusted PE Ratio 220 -120 012 750 -140
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 25 36
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Examples
Whatrsquos next
Constrained estimation of Θ (Linear constraints rank constraints)
Generalize to higher dimensions
Fancier hedging model
Conditional covariance models
Jakrsquos Tap
Thank You
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 26 36
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Appendix
Bibliography I
[1] Clifford S Asness Andrea Frazzini Ronen Israel and Tobias J Moskowitz Fact fiction and momentum investingPrivately Published May 2014 URL httpssrncomabstract=2435323
[2] Taras Bodnar and Yarema Okhrin On the product of inverse Wishart and normal distributions with applications todiscriminant analysis and portfolio theory Scandinavian Journal of Statistics 38(2)311ndash331 2011 ISSN 1467-9469 doi101111j1467-9469201100729x URL httpdxdoiorg101111j1467-9469201100729x
[3] Michael W Brandt Portfolio choice problems Handbook of financial econometrics 1269ndash336 2009 URLhttpshrreceptidocsrudocs54748conv_1file1pdfpage=298
[4] Mark Britten-Jones The sampling error in estimates of mean-variance efficient portfolio weights The Journal of Finance54(2)655ndash671 1999 URL httpwwwjstororgstable2697722
[5] Vijay Kumar Chopra and William T Ziemba The effect of errors in means variances and covariances on optimal portfoliochoice The Journal of Portfolio Management 19(2)6ndash11 1993 URLhttpfacultyfuquadukeedu~charveyTeachingBA453_2006Chopra_The_effect_of_1993pdf
[6] Gregory Connor Sensible return forecasting for portfolio management Financial Analysts Journal 53(5)pp 44ndash51 1997ISSN 0015198X URL https
facultyfuquadukeedu~charveyTeachingBA453_2006Connor_Sensible_Return_Forecasting_1997pdf
[7] Paul L Fackler Notes on matrix calculus Privately Published 2005 URLhttpwww4ncsuedu~pfacklerMatCalcpdf
[8] Narayan C Giri On the likelihood ratio test of a normal multivariate testing problem The Annals of MathematicalStatistics 35(1)181ndash189 1964 doi 101214aoms1177703740 URLhttpprojecteuclidorgeuclidaoms1177703740
[9] Ulf Herold and Raimond Maurer Tactical asset allocation and estimation risk Financial Markets and PortfolioManagement 18(1)39ndash57 2004 ISSN 1555-4961 doi 101007s11408-004-0104-2 URLhttpdxdoiorg101007s11408-004-0104-2
[10] Gur Huberman and Shmuel Kandel Mean-variance spanning The Journal of Finance 42(4)pp 873ndash888 1987 ISSN00221082 URL httpwwwjstororgstable2328296
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 27 36
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Appendix
Bibliography II
[11] Raymond Kan and GuoFu Zhou Tests of mean-variance spanning Annals of Economics and Finance 13(1) 2012 URLhttpwwwaeconfnetArticlesMay2012aef130105pdf
[12] Jan R Magnus and H Neudecker Matrix Differential Calculus with Applications in Statistics and Econometrics WileySeries in Probability and Statistics Texts and References Section Wiley 3rd edition 2007 ISBN 9780471986331 URLhttpwwwjanmagnusnlmiscmdc2007-3rdedition
[13] Raymond McTaggart and Gergely Daroczi Quandl Quandl Data Connection 2014 URLhttpCRANR-projectorgpackage=Quandl R package version 232
[14] Yarema Okhrin and Wolfgang Schmid Distributional properties of portfolio weights Journal of Econometrics 134(1)235ndash256 2006 URL httpwwwsciencedirectcomsciencearticlepiiS0304407605001442
[15] Steven E Pav MarkowitzR Statistical significance of the Markowitz portfolio 2014 URLhttpwwwr-projectorghttpsgithubcomshabbychefMarkowitzR R package version 01403
[16] C Radhakrishna Rao Advanced Statistical Methods in Biometric Research John Wiley and Sons 1952 URLhttpbooksgooglecombooksid=HvFLAAAAMAAJ
[17] A D Roy Safety first and the holding of assets Econometrica 20(3)pp 431ndash449 1952 ISSN 00129682 URLhttpwwwjstororgstable1907413
[18] Mervyn J Silvapulle and Pranab Kumar Sen Constrained statistical inference inequality order and shape restrictionsWiley-Interscience Hoboken NJ 2005 ISBN 0471208272 URL httpbooksgooglecombooksisbn=0471208272
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 28 36
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Appendix
Common Questions I
Doesnrsquot this require fourth order moments
I always use relative (or lsquopercentrsquo) returns These are bounded Allmoments exist Identical distribution is a much more questionableassumption
Isnrsquot the complexity Ω(p4)
Portfolio optimization for large p (bigger than 20) is not typicallyrecommended
Wonrsquot estimating a large number of parameters hurt performance
The covariance Var(vech
(xxgt
))has Ω
(p4)
elements but the portfolio isconstructed only from Ω
(p2)
elements as with vanilla Markowitz
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 29 36
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Appendix
Common Questions II
I want to hedge out exposure to a non-asset
I want that as well It does not appear to be a simple modification of theweird trick but it may be one discovery away
I want to maximize Sharpe ratio with a time-dependent risk-free rate
I suspect that the lsquorightrsquo way to do this is to include the RFR as an assetthen hedge out exposure to it This effectively allows each asset to have anon-unit lsquobetarsquo to the risk-free which seems like a higher bar than justhedging a constant unit of the risk-free
What was the quote about the pterodactyl
It was from the movie Airplane
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 30 36
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Appendix
Common Questions III
I want to hedge out an asset but I do not want the mean of that asset tobe estimated
I believe this can be done with constrained estimation of Θ Briefly if thereare linear constraints one believes Θ satisfies you can solve a least-squaresproblem to get a sample estimate which satisfies the constraints and is nottoo lsquofarrsquo from the unconstrained estimator I have not done the analysisbut believe it is another simple application of the delta method
The conditional expectation model is many-to-many How do I sparseify it
Similar to the above but I believe one would want to specify linearconstraints on the Cholesky factor of Θ This might be more complicatedOr maybe not
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 31 36
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Appendix
Common Questions IV
I donrsquot want to deal with the headaches of symmetry
The Cholesky factor of Θ is
[1 0
micro Σ12
] This is a lower triangular
matrix and completely determines Θ I suspect much of the analysis canbe re-couched in terms of this square root but I do not know the matrixderivative of the Cholesky factorization
What about a mashup with Kalman Filters
Sure This should probably be expressed as an update on the Choleskyfactor Θ12
Which portfolio managers are using the weird trick
All of them except you
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 32 36
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Appendix
Common Questions V
I am not comforted by the fact that ζ2lowast ζ2
lowast since the portfolio νlowast mayachieve a much lower Sharpe ratio than optimal
Because νlowast is the optimal population Sharpe ratio of any portfolio it is anupper bound on the Sharpe ratio of νlowast To estimate the lsquogaprsquo requires Ibelieve the second-order multivariate delta method I have not done theanalysis
Can you shoehorn a short-sale constraint into the model
I doubt it is feasible It is known for example that Hotellingrsquos statisticunder a positivity constraint is not a similar statistic indicating Sharperatio is an imperfect yardstick for sign-constrained portfolio problems [18]
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 33 36
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Appendix
Common Questions VI
Why maximize Sharpe ratio Everyone else maximizes lsquoutilityrsquo
No investor has ever told us their lsquorisk aversion parameterrsquo but they askabout our Sharpe ratio all the time Also read Roy for the connectionbetween Sharpe ratio and probability of a loss [17]
How do you deal with trade costs
It is not clear One hack would be to assume trade costs quadratic in thetarget portfolio I believe this merely leads to an inflation of the Σ butthere are likely complications
Isnrsquot independence of ˜xi suspicious
If the state variables wi depend on the previous period returns xi independence will be violated However the CLT may apply if thesequence is weakly dependent or lsquostrongly mixingrsquo
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 34 36
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Appendix
Common Questions VII
How do you detect outliers
This probably requires one to impose a likelihood on ˜xi
Does the math simplify if you assume normal returns
In this case nΘ takes a conditional Wishart distribution
But does it do big data
Computation of Θ is very simple since it is just an uncentered moment
How should a Bayesian approach estimation of Θ
I donrsquot know Ask one I suspect they would assume normal returns thenassume some kind of conditional Wishart prior
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 35 36
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-
Appendix
Common Questions VIII
Does the hedged portfolio involve a projection
It does The hedged portfolio is the optimal portfolio minus a projectionunder the metric induced by Σ
It seems that when I hedge out a single asset only the holdings in thatasset change in the portfolio
If you look at the projection operation the change can only occur in thecolumn space of Ggt which in this case means only the holdings in thesingle asset will change (This is all modulo adjustments to overall grossleverage to meet the risk budget)
Can you back out the traditional significance tests from the asymptoticdistribution of Θ
Possibly but probably a bit uglier than I can stomach
Steven Pav (Cerebellum Capital) Portfolio Inference May 16 2014 36 36
- One Weird Trick
- Portfolio Inference
- Extensions
- Examples
- Appendix
-