CHAPTER EIGHT

Case #16: SEASONAL AIR TRAVEL AND FORECAST COMBINATION

Goal: This utilizes two univariate time-series models to produce forecasts and examines whether different ways to model the same information justify the efforts involved in forecast combination.

Specifically, this case examines:

Problem Spreadsheet

The spreadsheet for this problem is C8_Case2.xls. It contains revenue passenger miles (RPM) data for airline travel.

Variable

Data Range

RPM

1979M1-1984M2

Data on international revenue-passenger miles (RPM) is published monthly by the U.S. Department of Transportation. (Recall this data was analyzed in Problem #3 in this chapter).

Our goal is to examine improvement in model fit using forecast combination over the sample period.

Examining Data for Stationarity

To examine graphically any trend or seasonality in the data, we generated a time-series plot of the RPM data.

Question #1: Based on examination of the time-series plot of RPM, does the data exhibit significant trend or seasonality?

ANSWER:

To further analyze the data for seasonality, generate ACF and PACF correlograms reported below.

Question #2: Are the data stationary? Explain.

ANSWER:

Accordingly, for this case we have seasonal data with a trend. To handle these complications requires some form of data transformation and/or modeling of seasonality and trend. To accomplish this we shall employ the Expert Selection aspects of FORECASTXTM, and then combine forecasts.

Box-Jenkins Forecasts

Using FORECASTXTM, we first developed a Box-Jenkins model using historical data with no holdback or forecast period designated.

Estimation results are presented below.

Method Selected

Box Jenkins

Model Selected

ARIMA(0,1,0) * (1,0,0)

T-Test For Seasonal AR

9.53

Note that FORECASTXTM chooses an ARIMA(0,1,0)*(1,0,0) model as optimal.

In-sample accuracy statistics are reported below.

Accuracy Measures

 

Value

AIC

492.82

BIC

494.94

Mean Absolute Percentage Error (MAPE)

0.42%

Sum Squared Error (SSE)

9,953.04

R-Square

99.83%

Adjusted R-Square

99.83%

Chi-Square

0.00

Cochrane-Orcutt

0.31

Mean Absolute Error

9.87

Mean Error

3.80

Mean Square Error

160.53

Normality Error

1.82

Root Mean Square Error

12.67

Standard Deviation of Error

12.77

Theil

0.64

Question #3: How well does our ARIMA(0,1,0)*(1,0,0) model fit the data?

ANSWER:

Winter’s Smoothing Forecasts

An alternative way of forecasting RPM is Winter’s smoothing which handles trend and seasonality.

Estimation results using Winter’s smoothing are reported below.

Method Statistics

 

Value

Method Selected

Double Exponential Smoothing-Holt

Alpha

1.00

Gamma

0.17

Accuracy Measures

 

Value

AIC

463.89

BIC

468.14

Mean Absolute Percentage Error (MAPE)

0.32%

Sum Squared Error (SSE)

6,043.95

R-Square

99.90%

Adjusted R-Square

99.90%

Chi-Square

0.00

Cochrane-Orcutt

0.08

Mean Absolute Error

7.45

Mean Error

0.64

Mean Square Error

97.48

Normality Error

1.82

Root Mean Square Error

9.87

Standard Deviation of Error

9.95

Theil

0.49

Question #4: How well does the Winter’s model fit the data?

ANSWER:

Question #5: Can you make the case that these two methods should be combined to generate forecasts?

ANSWER:

Can Our Forecasts be Combined?

To examine whether forecast combination is possible, we estimate the following regression over the sample period.

Audit Trail -- Coefficient Table (Multiple Regression Selected)

 

 

Series

Included

Standard

Overall

Description

in Model

Coefficient

Error

T-test

F-test

RPM

Dependent

-5.12

10.06

-0.51

28,958.15

ARIMA_Forecast

Yes

0.02

0.18

0.13

Winters_Forecast

Yes

0.98

0.18

5.51

 

Question #6: Can the two models be combined to generate unbiased forecasts? Explain.

ANSWER:

Optimal Forecast Combination

Next, calculate the optimal weights to be given to each method by estimating the following regression:

Multiple Regression -- Result Formula

 

 

RPM = 0. + ( (ARIMA_Forecast) * 0.034396 ) + ( (Winters_Forecast) * 0.965955 )

Audit Trail -- Coefficient Table (Multiple Regression Selected)

 

 

Series

Included

Standard

Overall

Description

in Model

Coefficient

Error

T-test

F-test

RPM

Dependent

0.00

0.00

0.00

29,320.17

ARIMA_Forecast

Yes

0.03

0.17

0.20

Winters_Forecast

Yes

0.97

0.17

5.53

 

In-sample fit statistics are reported below.

Accuracy Measures

 

Value

AIC

461.50

BIC

463.63

Mean Absolute Percentage Error (MAPE)

0.32%

Sum Squared Error (SSE)

6,006.46

R-Square

99.90%

Adjusted R-Square

99.89%

Chi-Square

0.00

Cochrane-Orcutt

0.08

Mean Absolute Error

7.48

Mean Error

-0.08

Mean Square Error

96.88

Normality Error

1.82

Root Mean Square Error

9.84

Standard Deviation of Error

9.92

Theil

0.48

An in-sample plot of RPM and combined model forecasts is reported below.

Question #7: Report the optimal weights below. Explain!

ANSWER:

In-sample RMSE for a three models is reported below.

Model

In-Sample RMSE

ARIMA(0,1,0)*(1,0,0)

12.67

Winter’s Smoothing

9.87

Combined Model

9.84

 

Question #8: Based upon the combined forecasts results reported above, are there benefits to combining these two forecasting models?

ANSWER:

Student Practice Question

Question #1: Redo this case with another variable that is highly seasonal. Contrast and compare your results with those of this case.