Forecasting: principles and practice · Estimation If we minimize Pn2 t (by using ordinary...
56
Forecasting: principles and practice 1 Forecasting: principles and practice Rob J Hyndman 3.2 Dynamic regression
Transcript of Forecasting: principles and practice · Estimation If we minimize Pn2 t (by using ordinary...
Forecasting: principles and practice 1
Forecasting: principlesand practice
Rob J Hyndman
3.2 Dynamic regression
Outline
1 Regression with ARIMA errors
2 Stochastic and deterministic trends
3 Periodic seasonality
4 Lab session 14
5 Dynamic regression models
Forecasting: principles and practice Regression with ARIMA errors 2
Regression with ARIMA errorsRegression models
yt = β0 + β1x1,t + · · · + βkxk,t + et,
yt modeled as function of k explanatory variablesx1,t, . . . , xk,t.In regression, we assume that et was WN.Now we want to allow et to be autocorrelated.
Example: ARIMA(1,1,1) errorsyt = β0 + β1x1,t + · · · + βkxk,t + nt,(1− φ1B)(1− B)nt = (1 + θ1B)et,
where et is white noise .Forecasting: principles and practice Regression with ARIMA errors 3
Regression with ARIMA errorsRegression models
yt = β0 + β1x1,t + · · · + βkxk,t + et,
yt modeled as function of k explanatory variablesx1,t, . . . , xk,t.In regression, we assume that et was WN.Now we want to allow et to be autocorrelated.
Example: ARIMA(1,1,1) errorsyt = β0 + β1x1,t + · · · + βkxk,t + nt,(1− φ1B)(1− B)nt = (1 + θ1B)et,
where et is white noise .Forecasting: principles and practice Regression with ARIMA errors 3
Residuals and errors
Example: nt = ARIMA(1,1,1)yt = β0 + β1x1,t + · · · + βkxk,t + nt,(1− φ1B)(1− B)nt = (1 + θ1B)et,
Be careful in distinguishing nt from et.Only the errors nt are assumed to be white noise.In ordinary regression, nt is assumed to be whitenoise and so nt = et.
Forecasting: principles and practice Regression with ARIMA errors 4
Residuals and errors
Example: nt = ARIMA(1,1,1)yt = β0 + β1x1,t + · · · + βkxk,t + nt,(1− φ1B)(1− B)nt = (1 + θ1B)et,
Be careful in distinguishing nt from et.Only the errors nt are assumed to be white noise.In ordinary regression, nt is assumed to be whitenoise and so nt = et.
Forecasting: principles and practice Regression with ARIMA errors 4
Estimation
If we minimize ∑ n2t (by using ordinary regression):
1 Estimated coefficients β̂0, . . . , β̂k are no longeroptimal as some information ignored;
2 Statistical tests associated with the model (e.g.,t-tests on the coefficients) are incorrect.
3 p-values for coefficients usually too small (“spuriousregression”).
4 AIC of fitted models misleading.
Minimizing ∑ e2t avoids these problems.Maximizing likelihood is similar to minimizing ∑ e2t .
Forecasting: principles and practice Regression with ARIMA errors 5
Estimation
If we minimize ∑ n2t (by using ordinary regression):
1 Estimated coefficients β̂0, . . . , β̂k are no longeroptimal as some information ignored;
2 Statistical tests associated with the model (e.g.,t-tests on the coefficients) are incorrect.
3 p-values for coefficients usually too small (“spuriousregression”).
4 AIC of fitted models misleading.
Minimizing ∑ e2t avoids these problems.Maximizing likelihood is similar to minimizing ∑ e2t .
Forecasting: principles and practice Regression with ARIMA errors 5
Stationarity
Model with ARIMA(1,1,1) errorsyt = β0 + β1x1,t + · · · + βkxk,t + nt,(1− φ1B)(1− B)nt = (1 + θ1B)et,
Equivalent to model with ARIMA(1,0,1) errorsy′t = β1x
′1,t + · · · + βkx′k,t + n′t,
(1− φ1B)n′t = (1 + θ1B)et,
where y′t = yt − yt−1, x′t,i = xt,i − xt−1,i and n′t = nt − nt−1.
Forecasting: principles and practice Regression with ARIMA errors 6
Stationarity
Model with ARIMA(1,1,1) errorsyt = β0 + β1x1,t + · · · + βkxk,t + nt,(1− φ1B)(1− B)nt = (1 + θ1B)et,
Equivalent to model with ARIMA(1,0,1) errorsy′t = β1x
′1,t + · · · + βkx′k,t + n′t,
(1− φ1B)n′t = (1 + θ1B)et,
where y′t = yt − yt−1, x′t,i = xt,i − xt−1,i and n′t = nt − nt−1.
Forecasting: principles and practice Regression with ARIMA errors 6
Regression with ARIMA errorsAny regression with an ARIMA error can be rewritten as aregression with an ARMA error by differencing all variableswith the same differencing operator as in the ARIMAmodel.Original data
yt = β0 + β1x1,t + · · · + βkxk,t + ntwhere φ(B)(1− B)dnt = θ(B)et
After differencing all variablesy′t = β1x
′1,t + · · · + βkx′k,t + n′t.
where φ(B)nt = θ(B)etand y′t = (1− B)dyt
Forecasting: principles and practice Regression with ARIMA errors 7
Regression with ARIMA errorsAny regression with an ARIMA error can be rewritten as aregression with an ARMA error by differencing all variableswith the same differencing operator as in the ARIMAmodel.Original data
yt = β0 + β1x1,t + · · · + βkxk,t + ntwhere φ(B)(1− B)dnt = θ(B)et
After differencing all variablesy′t = β1x
′1,t + · · · + βkx′k,t + n′t.
where φ(B)nt = θ(B)etand y′t = (1− B)dyt
Forecasting: principles and practice Regression with ARIMA errors 7
Regression with ARIMA errorsAny regression with an ARIMA error can be rewritten as aregression with an ARMA error by differencing all variableswith the same differencing operator as in the ARIMAmodel.Original data
yt = β0 + β1x1,t + · · · + βkxk,t + ntwhere φ(B)(1− B)dnt = θ(B)et
After differencing all variablesy′t = β1x
′1,t + · · · + βkx′k,t + n′t.
where φ(B)nt = θ(B)etand y′t = (1− B)dyt
Forecasting: principles and practice Regression with ARIMA errors 7
Model selection
Fit regression model with automatically selectedARIMA errors.Check that et series looks like white noise.Note that estimation is done on the differenced seriesto ensure consistent estimators.
Selecting predictorsAICc can be calculated for final model.Repeat procedure for all subsets of predictors to beconsidered, and select model with lowest AICc value.
Forecasting: principles and practice Regression with ARIMA errors 8
US personal consumption & incomeC
onsumption
Income
Production
Savings
Unem
ployment
1970 1980 1990 2000 2010
−2−1
012
−2.50.02.5
−5.0−2.5
0.02.5
−50−25
02550
−1.0−0.5
0.00.51.01.5
Year
Quarterly changes in US consumption and personal income
Forecasting: principles and practice Regression with ARIMA errors 9
US personal consumption & income
−2
−1
0
1
2
−2.5 0.0 2.5
Income
Con
sum
ptio
n
Quarterly changes in US consumption and personal income
Forecasting: principles and practice Regression with ARIMA errors 10
US Personal Consumption and income
No need for transformations or further differencing.Increase in income does not necessarily translate intoinstant increase in consumption (e.g., after the loss ofa job, it may take a few months for expenses to bereduced to allow for the new circumstances). We willignore this for now.
Forecasting: principles and practice Regression with ARIMA errors 11
US personal consumption & income
(fit <- auto.arima(uschange[,"Consumption"],xreg=uschange[,"Income"]))
## Series: uschange[, "Consumption"]## Regression with ARIMA(1,0,2) errors#### Coefficients:## ar1 ma1 ma2 intercept xreg## 0.6922 -0.5758 0.1984 0.5990 0.2028## s.e. 0.1159 0.1301 0.0756 0.0884 0.0461#### sigma^2 estimated as 0.3219: log likelihood=-156.95## AIC=325.91 AICc=326.37 BIC=345.29
Forecasting: principles and practice Regression with ARIMA errors 12
US personal consumption & income
ggtsdisplay(residuals(fit, type='regression'),main="ARIMA errors")
−2
−1
0
1
1970 1980 1990 2000 2010
ARIMA errors
−0.1
0.0
0.1
0.2
0.3
4 8 12 16 20
Lag
AC
F
−0.1
0.0
0.1
0.2
0.3
4 8 12 16 20
Lag
PAC
F
Forecasting: principles and practice Regression with ARIMA errors 13
US personal consumption & income
ggtsdisplay(residuals(fit),main="ARIMA residuals")
−2
−1
0
1
1970 1980 1990 2000 2010
ARIMA residuals
−0.1
0.0
0.1
4 8 12 16 20
Lag
AC
F
−0.1
0.0
0.1
4 8 12 16 20
Lag
PAC
F
Forecasting: principles and practice Regression with ARIMA errors 14
US Personal Consumption and Income
A Ljung-Box test shows the residuals are uncorrelated.
checkresiduals(fit, plot=FALSE)
#### Ljung-Box test#### data: Residuals from Regression with ARIMA(1,0,2) errors## Q* = 5.8916, df = 3, p-value = 0.117#### Model df: 5. Total lags used: 8
Forecasting: principles and practice Regression with ARIMA errors 15
US Personal Consumption and Income
fcast <- forecast(fit,xreg=rep(mean(uschange[,"Income"]),8), h=8)
autoplot(fcast) + xlab("Year") +ylab("Percentage change") +ggtitle("Forecasts from regression with ARIMA(1,0,2) errors")
−2
−1
0
1
2
1970 1980 1990 2000 2010 2020
Year
Per
cent
age
chan
ge
level
80
95
Forecasts from regression with ARIMA(1,0,2) errors
Forecasting: principles and practice Regression with ARIMA errors 16
Forecasting
To forecast a regression model with ARIMA errors, weneed to forecast the regression part of the model andthe ARIMA part of the model and combine the results.
Some explanatory variable are known into the future (e.g.,time, dummies).Separate forecasting models may be needed for otherexplanatory variables.
Forecasting: principles and practice Regression with ARIMA errors 17
Outline
1 Regression with ARIMA errors
2 Stochastic and deterministic trends
3 Periodic seasonality
4 Lab session 14
5 Dynamic regression models
Forecasting: principles and practice Stochastic and deterministic trends 18
Stochastic & deterministic trends
Deterministic trend
yt = β0 + β1t + nt
where nt is ARMA process.Stochastic trend
yt = β0 + β1t + nt
where nt is ARIMA process with d ≥ 1.Difference both sides until nt is stationary:
y′t = β1 + n′t
where n′t is ARMA process.Forecasting: principles and practice Stochastic and deterministic trends 19
Stochastic & deterministic trends
Deterministic trend
yt = β0 + β1t + nt
where nt is ARMA process.Stochastic trend
yt = β0 + β1t + nt
where nt is ARIMA process with d ≥ 1.Difference both sides until nt is stationary:
y′t = β1 + n′t
where n′t is ARMA process.Forecasting: principles and practice Stochastic and deterministic trends 19
Stochastic & deterministic trends
Deterministic trend
yt = β0 + β1t + nt
where nt is ARMA process.Stochastic trend
yt = β0 + β1t + nt
where nt is ARIMA process with d ≥ 1.Difference both sides until nt is stationary:
y′t = β1 + n′t
where n′t is ARMA process.Forecasting: principles and practice Stochastic and deterministic trends 19
International visitors
2
4
6
1980 1990 2000 2010
Year
mill
ions
of p
eopl
e
Total annual international visitors to Australia
Forecasting: principles and practice Stochastic and deterministic trends 20
International visitorsDeterministic trend
(fit1 <- auto.arima(austa, d=0, xreg=1:length(austa)))
## Series: austa## Regression with ARIMA(2,0,0) errors#### Coefficients:## ar1 ar2 intercept xreg## 1.1127 -0.3805 0.4156 0.1710## s.e. 0.1600 0.1585 0.1897 0.0088#### sigma^2 estimated as 0.02979: log likelihood=13.6## AIC=-17.2 AICc=-15.2 BIC=-9.28
yt = 0.4173 + 0.1715t + ntnt = 1.0371nt−1 − 0.3379nt−2 + etet ∼ NID(0, 0.02854).
Forecasting: principles and practice Stochastic and deterministic trends 21
International visitorsDeterministic trend
(fit1 <- auto.arima(austa, d=0, xreg=1:length(austa)))
## Series: austa## Regression with ARIMA(2,0,0) errors#### Coefficients:## ar1 ar2 intercept xreg## 1.1127 -0.3805 0.4156 0.1710## s.e. 0.1600 0.1585 0.1897 0.0088#### sigma^2 estimated as 0.02979: log likelihood=13.6## AIC=-17.2 AICc=-15.2 BIC=-9.28
yt = 0.4173 + 0.1715t + ntnt = 1.0371nt−1 − 0.3379nt−2 + etet ∼ NID(0, 0.02854).
Forecasting: principles and practice Stochastic and deterministic trends 21
International visitorsStochastic trend
(fit2 <- auto.arima(austa,d=1))
## Series: austa## ARIMA(0,1,1) with drift#### Coefficients:## ma1 drift## 0.3006 0.1735## s.e. 0.1647 0.0390#### sigma^2 estimated as 0.03376: log likelihood=10.62## AIC=-15.24 AICc=-14.46 BIC=-10.57
yt − yt−1 = 0.1537 + etyt = y0 + 0.1537t + ntnt = nt−1 + etet ∼ NID(0, 0.03241).Forecasting: principles and practice Stochastic and deterministic trends 22
International visitorsStochastic trend
(fit2 <- auto.arima(austa,d=1))
## Series: austa## ARIMA(0,1,1) with drift#### Coefficients:## ma1 drift## 0.3006 0.1735## s.e. 0.1647 0.0390#### sigma^2 estimated as 0.03376: log likelihood=10.62## AIC=-15.24 AICc=-14.46 BIC=-10.57
yt − yt−1 = 0.1537 + etyt = y0 + 0.1537t + ntnt = nt−1 + etet ∼ NID(0, 0.03241).Forecasting: principles and practice Stochastic and deterministic trends 22
International visitors
2.5
5.0
7.5
1980 1990 2000 2010 2020
Year
level
80
95
Forecasts from linear trend with AR(2) error
2.5
5.0
7.5
10.0
1980 1990 2000 2010 2020
Year
level
80
95
Forecasts from ARIMA(0,1,0) with drift
Forecasting: principles and practice Stochastic and deterministic trends 23
Forecasting with trend
Point forecasts are almost identical, but predictionintervals differ.Stochastic trends have much wider predictionintervals because the errors are non-stationary.Be careful of forecasting with deterministic trends toofar ahead.
Forecasting: principles and practice Stochastic and deterministic trends 24
Outline
1 Regression with ARIMA errors
2 Stochastic and deterministic trends
3 Periodic seasonality
4 Lab session 14
5 Dynamic regression models
Forecasting: principles and practice Periodic seasonality 25
Fourier terms for seasonalityPeriodic seasonality can be handled using pairs of Fourierterms:
sk(t) = sin(2πkt
m
)ck(t) = cos
(2πktm
)
yt =K∑k=1
[αksk(t) + βkck(t)] + nt
nt is non-seasonal ARIMA process.Every periodic function can be approximated by sumsof sin and cos terms for large enough K.Choose K by minimizing AICc.
Forecasting: principles and practice Periodic seasonality 26
US Accidental Deaths
fit <- auto.arima(USAccDeaths,xreg=fourier(USAccDeaths, 5),seasonal=FALSE)
fc <- forecast(fit,xreg=fourier(USAccDeaths, 5, 24))
Forecasting: principles and practice Periodic seasonality 27
US Accidental Deaths
autoplot(fc)
8000
10000
12000
1973 1975 1977 1979 1981
Time
US
Acc
Dea
ths
level
80
95
Forecasts from Regression with ARIMA(0,1,1) errors
Forecasting: principles and practice Periodic seasonality 28
Outline
1 Regression with ARIMA errors
2 Stochastic and deterministic trends
3 Periodic seasonality
4 Lab session 14
5 Dynamic regression models
Forecasting: principles and practice Lab session 14 29
Lab Session 14
Forecasting: principles and practice Lab session 14 30
Outline
1 Regression with ARIMA errors
2 Stochastic and deterministic trends
3 Periodic seasonality
4 Lab session 14
5 Dynamic regression models
Forecasting: principles and practice Dynamic regression models 31
Dynamic regression models
Sometimes a change in xt does not affect ytinstantaneously
yt = sales, xt = advertising.yt = stream flow, xt = rainfall.yt = size of herd, xt = breeding stock.
These are dynamic systems with input (xt) and output(yt).xt is often a leading indicator.There can be multiple predictors.
Forecasting: principles and practice Dynamic regression models 32
Dynamic regression models
Sometimes a change in xt does not affect ytinstantaneously
yt = sales, xt = advertising.yt = stream flow, xt = rainfall.yt = size of herd, xt = breeding stock.
These are dynamic systems with input (xt) and output(yt).xt is often a leading indicator.There can be multiple predictors.
Forecasting: principles and practice Dynamic regression models 32
Dynamic regression models
Sometimes a change in xt does not affect ytinstantaneously
yt = sales, xt = advertising.yt = stream flow, xt = rainfall.yt = size of herd, xt = breeding stock.
These are dynamic systems with input (xt) and output(yt).xt is often a leading indicator.There can be multiple predictors.
Forecasting: principles and practice Dynamic regression models 32
Lagged explanatory variablesThe model include present and past values of predictor:xt, xt−1, xt−2, . . . .
yt = a + ν0xt + ν1xt−1 + · · · + νkxt−k + nt
where nt is an ARIMA process.Rewrite model as
yt = a + (ν0 + ν1B + ν2B2 + · · · + νkBk)xt + nt= a + ν(B)xt + nt.
ν(B) is called a transfer function since it describeshow change in xt is transferred to yt.x can influence y, but y is not allowed to influence x.Forecasting: principles and practice Dynamic regression models 33
Lagged explanatory variablesThe model include present and past values of predictor:xt, xt−1, xt−2, . . . .
yt = a + ν0xt + ν1xt−1 + · · · + νkxt−k + nt
where nt is an ARIMA process.Rewrite model as
yt = a + (ν0 + ν1B + ν2B2 + · · · + νkBk)xt + nt= a + ν(B)xt + nt.
ν(B) is called a transfer function since it describeshow change in xt is transferred to yt.x can influence y, but y is not allowed to influence x.Forecasting: principles and practice Dynamic regression models 33
Lagged explanatory variablesThe model include present and past values of predictor:xt, xt−1, xt−2, . . . .
yt = a + ν0xt + ν1xt−1 + · · · + νkxt−k + nt
where nt is an ARIMA process.Rewrite model as
yt = a + (ν0 + ν1B + ν2B2 + · · · + νkBk)xt + nt= a + ν(B)xt + nt.
ν(B) is called a transfer function since it describeshow change in xt is transferred to yt.x can influence y, but y is not allowed to influence x.Forecasting: principles and practice Dynamic regression models 33
Example: Insurance quotes and TV advertsQ
uotesT
V.advert
2002 2003 2004 2005
8
10
12
14
16
18
6
7
8
9
10
11
Year
Insurance advertising and quotations
Forecasting: principles and practice Dynamic regression models 34
Example: Insurance quotes and TV adverts
Advert <- cbind(insurance[,2], c(NA,insurance[1:39,2]))colnames(Advert) <- paste("AdLag",0:1,sep="")(fit <- auto.arima(insurance[,1], xreg=Advert, d=0))
## Series: insurance[, 1]## Regression with ARIMA(3,0,0) errors#### Coefficients:## ar1 ar2 ar3 intercept AdLag0 AdLag1## 1.4117 -0.9317 0.3591 2.0393 1.2564 0.1625## s.e. 0.1698 0.2545 0.1592 0.9931 0.0667 0.0591#### sigma^2 estimated as 0.2165: log likelihood=-23.89## AIC=61.78 AICc=65.28 BIC=73.6
yt = 2.04 + 1.26xt + 0.16xt−1 + ntnt = 1.41nt−1 − 0.93nt−2 + 0.36nt−3 + et
Forecasting: principles and practice Dynamic regression models 35
Example: Insurance quotes and TV adverts
Advert <- cbind(insurance[,2], c(NA,insurance[1:39,2]))colnames(Advert) <- paste("AdLag",0:1,sep="")(fit <- auto.arima(insurance[,1], xreg=Advert, d=0))
## Series: insurance[, 1]## Regression with ARIMA(3,0,0) errors#### Coefficients:## ar1 ar2 ar3 intercept AdLag0 AdLag1## 1.4117 -0.9317 0.3591 2.0393 1.2564 0.1625## s.e. 0.1698 0.2545 0.1592 0.9931 0.0667 0.0591#### sigma^2 estimated as 0.2165: log likelihood=-23.89## AIC=61.78 AICc=65.28 BIC=73.6
yt = 2.04 + 1.26xt + 0.16xt−1 + ntnt = 1.41nt−1 − 0.93nt−2 + 0.36nt−3 + et
Forecasting: principles and practice Dynamic regression models 35
Example: Insurance quotes and TV adverts
fc <- forecast(fit, h=20,xreg=cbind(c(Advert[40,1],rep(10,19)), rep(10,20)))
autoplot(fc)
8
10
12
14
16
18
2002 2003 2004 2005 2006 2007
Time
insu
ranc
e[, 1
]
level
80
95
Forecasts from Regression with ARIMA(3,0,0) errors
Forecasting: principles and practice Dynamic regression models 36
Example: Insurance quotes and TV adverts
fc <- forecast(fit, h=20,xreg=cbind(c(Advert[40,1],rep(8,19)), rep(8,20)))
autoplot(fc)
8
10
12
14
16
18
2002 2003 2004 2005 2006 2007
Time
insu
ranc
e[, 1
]
level
80
95
Forecasts from Regression with ARIMA(3,0,0) errors
Forecasting: principles and practice Dynamic regression models 37
Example: Insurance quotes and TV adverts
fc <- forecast(fit, h=20,xreg=cbind(c(Advert[40,1],rep(6,19)), rep(6,20)))
autoplot(fc)
8
10
12
14
16
18
2002 2003 2004 2005 2006 2007
Time
insu
ranc
e[, 1
]
level
80
95
Forecasts from Regression with ARIMA(3,0,0) errors
Forecasting: principles and practice Dynamic regression models 38
Dynamic regression modelsyt = a + ν(B)xt + nt
where nt is an ARMA process. So
φ(B)nt = θ(B)et or nt =θ(B)φ(B)
et = ψ(B)et.
yt = a + ν(B)xt + ψ(B)et
ARMA models are rational approximations to generaltransfer functions of et.We can also replace ν(B) by a rational approximation.There is no R package for forecasting using a generaltransfer function approach.Forecasting: principles and practice Dynamic regression models 39
Dynamic regression modelsyt = a + ν(B)xt + nt
where nt is an ARMA process. So
φ(B)nt = θ(B)et or nt =θ(B)φ(B)
et = ψ(B)et.
yt = a + ν(B)xt + ψ(B)et
ARMA models are rational approximations to generaltransfer functions of et.We can also replace ν(B) by a rational approximation.There is no R package for forecasting using a generaltransfer function approach.Forecasting: principles and practice Dynamic regression models 39
Dynamic regression modelsyt = a + ν(B)xt + nt
where nt is an ARMA process. So
φ(B)nt = θ(B)et or nt =θ(B)φ(B)
et = ψ(B)et.
yt = a + ν(B)xt + ψ(B)et
ARMA models are rational approximations to generaltransfer functions of et.We can also replace ν(B) by a rational approximation.There is no R package for forecasting using a generaltransfer function approach.Forecasting: principles and practice Dynamic regression models 39
