ETC3550 Applied Forecasting for Business and Economics 2018
Hello, dear friend, you can consult us at any time if you have any questions, add WeChat: daixieit
Semester One 2018
Examination Period
ETC3550
Applied Forecasting for Business and Economics
QUESTION 1
Write about a quarter of a page each on any FOUR of the following topics. (Clearly state if you agree or disagree with each statement. No marks will be given without any justification.)
(a) Prediction intervals are unnecessary because managers just want point forecasts.
(b) Whether we use a naïve approach, a decomposition, ETS or ARIMA models for forecasting, we always need to transform our data.
(c) Simple exponential smoothing should only be used for a series with a constant underlying level.
(d) All three information criteria: AIC, AICc and BIC are useful and they can potentially choose a diferent model. We prefer to use the AICc forecasting.
(e) The 95% prediction interval for an h-step-ahead naïve forecast is given by yˆT +h± 1.96′hσ2 .
(f) Regression models are not useful for forecasting because we always need to provide forecasts of the predictors.
QUESTION 2
Figure 1shows the total household expenditure for cigarette and tobacco consumption (CTC) in Victoria. The prices are represented as a chain volume measure (a representation of constant prices) in billions of dollars over the period 1985Q3–2017Q4.
Cigarettes and tobacco consumption total household expenditure for Victoria
2.5
2.0
1.5
2010
Cigarettes and tobacco total household expenditure for Victoria
2.5
2.0
1.5
Q1
Q2
Q3
Q4
(b) Do you think transforming this series will help? Why? Figure 3shows two possible transfor-
mations. Describe what you see and which one you would prefer to use.
(c) Figure 4 shows the forecasts generated from the stlf() function with the argument lambda=0. Describe what this function does and how the forecasts are generated.
2.5
2.0
1.5
1.0
2020
Time
Figure 4:
level
|
|
(d) You are asked to provide forecasts for the next two years for the CTC series shown in Figure 1. Consider applying each of the methods and models below. Comment, in a few words each, on whether each one is appropriate for forecasting the data. No marks will be given for simply guessing whether a method or a model is appropriate without justifying your choice.
A. Seasonal naïve method.
B. Drift method plus seasonal dummies.
C. Holt’s trend method.
D. Holt-Winters multiplicative damped trend method.
E. ETS(A,A,M).
F. ETS(M,Ad,M).
G. ARIMA(0,1,4).
H. ARIMA(3,1,2)(1,1,0)4 .
I. ARIMA(0,0,1)(2,0,0)4 .
J. Regression model with time and Fourier terms.
QUESTION 3
The following R code and output concerns two models for CTC (Cigarette and tobacco consumption) in Victoria.
fit1.ets <- ets (CTC)
fit1.ets
## ETS(M,N,M)
##
## Call:
## ets(y = CTC)
##
## Smoothing parameters:
## alpha = 0.709
## gamma = 0.291
##
## Initial states:
## l = 2.5802
## s=0.9818 0.9801 1.0598 0.9783
##
## sigma: 0.0308
##
## AIC AICc BIC
## -111.048 -110.130 -90.975
fit2.ets <- ets (CTC, lambda = 0 )
fit2.ets
## ETS(A,A,A)
##
## Call:
## ets(y = CTC, lambda = 0)
##
## Box-Cox transformation: lambda= 0
##
## Smoothing parameters:
## alpha = 0.5083
## beta = 1e-04
## gamma = 0.4536
##
## Initial states:
## l = 0.9583
## b = -0.0063
## s=-0.0017 -0.0154 0.056 -0.039
##
## sigma: 0.0309
##
## AIC AICc BIC
## -261.45 -259.95 -235.64
(a) Comment on the diferences between the two specifications, the models selected and the estimated parameters for each of the estimated models.
(b) Comment on Figure 5and how this relates to your answers above.
2.0
1.6
1.2 1.05
1.00
0.95
0.90 1.00
0.75
0.50
0.25
−0.00623 −0.00624 −0.00625 −0.00626 −0.00627 −0.00628
0.10
0.05
0.00
−0.05
2010
Time
Figure 5:
(c) Write down in full the estimated ETS model for fit1.ets.
(d) The output below and Figure 6relate to the residuals from fit1.ets. Comment on these in relation to the fit of the model. Give as many details as you can. What do your conclusions
here imply about using the model for forecasting?
Figure 6:
##
## Ljung-Box test
##
## data: Residuals from ETS(M,N,M)
## Q* = 14.7, df = 3, p-value = 0.0021
##
## Model df: 6. Total lags used: 9
(e) Use the output below to generate 2, 4 and 8-step-ahead forecasts. In row t, l=et , s1=st , . . . , s4=st-3 .
fit1.ets[["states"]] %>% tail()
## ## 2016 ## 2016 ## 2017 ## 2017 ## 2017 ## 2017 |
Q3 Q4 Q1 Q2 Q3 Q4 |
l 1.2802 1.2613 1.2707 1.2755 1.2756 1.2404 |
s1 0.92029 0.96256 0.92862 0.91493 0.92033 0.95164 |
s2 0.91352 0.92029 0.96256 0.92862 0.91493 0.92033 |
s3 0.92579 0.91352 0.92029 0.96256 0.92862 0.91493 |
s4 0.96841 0.92579 0.91352 0.92029 0.96256 0.92862
|
(f) Forecasts can be generated using fit2.ets %>% forecast(biasadj=TRUE). Comment on how biasadj=TRUE may be useful for this model.
QUESTION 4
(a) Figure 7shows four sets of ACFs and PACFs for the CTC (cigarettes and tobacco consumption) series of Victoria plotted in Figure 1. Explain what each of these show about the stationarity, seasonality and other features of the time series.
Figure 7:
(b) Use the relevant ACFs and PACFs from Figure 7to specify an appropriate ARIMA model with both seasonal and first diferencing.
(c) The following R code and output concerns a model for the CTC series. Write down the estimated model using backshift notation and expand this to the point where it can be used
to generate point forecasts.
auto.arima (CTC, lambda = 0 )
## Series: CTC
## ARIMA(1,0,0)(1,1,0)[4] with drift
## Box Cox transformation: lambda= 0
##
## Coefficients:
## ar1 sar1 drift
## 0.766 -0.378 -0.006
## s.e. 0.060 0.084 0.002
##
## sigma^2 estimated as 0.000912: log likelihood=263.13
## AIC=-518.26 AICc=-517.93 BIC=-506.91
(d) The last few values of the series are
## Qtr1 Qtr2 Qtr3 Qtr4
## 2015 1.248 1.229 1.253 1.263
## 2016 1.207 1.191 1.178 1.214
## 2017 1.180 1.167 1.174 1.180
Use the above model to calculate a forecast and a 95% prediction interval for 2018 Q1.
(e) The autoregressive coefcient of an AR(1) model yt = φyt-1 + ε T where εt ~ N (0, σ2 ) needs to be between -1 and 1 for the process to be stationary. Is this statement true or false. Justify your answer.
QUESTION 5
In the following code, a series of dynamic regression models are fitted to the CTC data. tt <- time (CTC)
t.break1 <- 1990
t.break2 <- 1996
t.break3 <- 2000
t1 <- pmax(0 ,tt -t.break1)
t2 <- pmax(0 ,tt -t.break2)
t3 <- pmax(0 ,tt -t.break3)
X123 <- cbind (tt,t1,t2,t3)
X23 <- cbind (tt,t2,t3)
X13 <- cbind (tt,t1,t3)
X12 <- cbind (tt,t1,t2)
X1 <- cbind (tt,t1)
X2 <- cbind (tt,t2)
X3 <- cbind (tt,t3)
X <- cbind(tt)
fit123 <- auto.arima (CTC, xreg=X123, lambda=0 )
fit23 <- auto.arima (CTC, xreg=X23, lambda=0 )
fit13 <- auto.arima (CTC, xreg=X13, lambda=0 )
fit12 <- auto.arima (CTC, xreg=X12, lambda=0 )
fit1 <- auto.arima (CTC, xreg=X1, lambda=0 )
fit2 <- auto.arima (CTC, xreg=X2, lambda=0 )
fit3 <- auto.arima (CTC, xreg=X3, lambda=0 )
fit <- auto.arima (CTC, xreg=X, lambda=0 )
AIC(fit123),AIC(fit23),AIC(fit13),AIC(fit12),
AIC(fit1),AIC(fit2),AIC(fit3),AIC(fit))
## [1] -551.66 -531.59 -532.36 -544.69 -533.20 -530.12 -528.20 -530.20
## Series: CTC
## Regression with ARIMA(1,0,0)(1,0,1)[4] errors
## Box Cox transformation: lambda= 0
##
## Coefficients:
## ar1 sar1 sma1 tt t1 t2 t3
## 0.5466 0.9546 -0.4459 5e-04 -0.0701 0.0806 -0.0364
## s.e. 0.0742 0.0230 0.0904 1e-04 0.0076 0.0133 0.0112
##
## sigma^2 estimated as 0.0007467: log likelihood=283.83
## AIC=-551.66 AICc=-550.47 BIC=-528.72
(a) Explain the process used in this code to select the final model, and write down the equations for the model.
(b) How could this approach be generalized to choose the number and position of knots automat-
ically? What problems do you imagine could happen when using such an algorithm?
(c) When producing forecasts using this model, what assumptions are you making about the piecewise linear trend? How does this compare with fitting a similar ETS model?
(d) Why is it better to use a seasonal ARIMA model here rather than Fourier terms to handle the seasonality in the data?
(e) Forecasts from the model are shown in Figure 8. Do you think they look reasonable? Explain.
Forecasts from Regression with ARIMA(1,0,0)(1,0,1)[4] errors
2.5
2.0
1.5
1.02020
Time
Figure 8:
level
|
|
2022-06-01