"Designated project" option

The following assigned project may be completed in lieu of an individually-designed project or the final exam. All projects are due by 5pm on the scheduled date of the final exam.

The data file called BA411PRO.SF (text version BA411PRO.TXT) consists of 30 times series downloaded from Datadisk for the period 1/72 to 5/96. (A few of the series start or end earlier than these dates.) Your mission: choose any two of the retail sales series, develop and compare several forecasting models for each, and ultimately come to a conclusion about what is the "best" model for forecasting each series one month in advance. You should compare and contrast the following kinds of models:

• ARIMA models
• exponential smoothing models, and
• regression models

and/or appropriate combinations of the above, with an appropriate number of points held out for validation.

Your analysis should, among other things, try to answer the questions:

• To what extent are inflation, real growth, seasonality, and cyclicality evident in the two chosen retail sales series? To what extent has the seasonal pattern (if any) been consistent over time?
• To what extent do "independent" variables (e.g., economic indicators) enable you to forecast the retail sales series better than you could forecast them from their own history alone--i.e., to what extent does a time series regression improve on a time series model without regressors? (Note: when using economic indicators as regressors, you should be careful to lag them by at least one month so that your model is capable of forecasting at least one month in advance.)
• What (if any) interesting recent trends do you notice, and what economic explanations (if any) can you find for them?
• Would you be confident in using any of your model(s) to try predict the series up to one or two years in advance?

Your writeup should include, as usual:

• appropriate model comparison reports (ideally located in the executive summary) ;
• calculations of "effective R-squared" (percent of variance explained beyond that of the best "naive" model--i.e., relative to the mean model fitted to a suitably differenced and/or logged version of the original series);
• an "audit trail" which includes an analysis summary report for each model that was compared, showing the exact variables(s) fitted, the sample sizes, and the estimated coefficients (note: relevant time series plots and auto- or cross-correlation plots should also be included, but you don't have to plot all 30 variables--just include appropriate plots of the variables that ended up being important in identifying and testing your models)
• the forecasts and confidence intervals produced by all models for the next period after the end of the data (this will be June 1996 in most cases);
• a discussion and comparison of the assumptions that underly each model;
• your "bottom line" conclusion about what is the best forecasting model for each of the two chosen retail sales series.

Variables in the BA411PRO data file:

CIBCRLONG     CIBCR long-leading composite index (1967=100)              

CIBCRSHORT    CIBCR short-leading composite index (1967=100)             

COINCIND      Index of 4 coincident indicators (1987=100)                

CONSCONF      Index of consumer confidence, The Conference Board         
              (1985=100)                                                 

CONSEXP       Index of  consumer expectations, The Conference Board      
              (1985=100)                                                 

CPI83         Consumer Price Index (1982-84=100)                         

DATE          Month/Year                                                 

GOLDPR        Gold price at New York; $/troy oz                          

HOUSTART      Housing starts, new private units (Thousands, ar)          

INDPROD       Industrial production index, total (1987=100)              

INDPRODCG     Industrial production index, consumer goods (1987=100)     

LAGIND        Index of seven lagging indicators (1987=100)               

LEADIND       Index of eleven leading indicators (1987=100)              

MFGINV        Manufacturer's inventories of finished goods ($ Bn)        

PERSINC       Personal income ($Bn, ar)                                  

RATIOIND      Ratio, coincident to lagging index (1987=100)              

RSAPPRAD      Retail sales, household appliance, ratio/TV/computer       
              stores, $Bn, nsa                                           

RSBOOK        Retail sales, book stores, $Bn, nsa                        

RSDRINK       Retail sales, drinking places (alcoholic beverages), $Bn,  
              nsa                                                        

RSEAT         Retail sales, eating places, $Bn, nsa                      

RSFURN        Retail sales, furniture stores, $Bn, nsa                   

RSJEWEL       Retail sales, jewelry stores, $Bn, nsa                     

RSLIQUOR      Retail sales, liquor stores, $Bn, nsa                      

RSSHOE        Retail sales, shoe stores, $Bn, nsa                        

RSSPORT       Retail sales, sporting goods stores and bicycle shops,     
              $Bn, nsa                                                   

RSWCLOTH      Retail sales, women's ready-to-wear clothing, $Bn, nsa     

SP500CLO      S&P stock price index, 500 composite, monthly close        

TBILL3MO      Treasury Bills, 3 month, auction rate average              

UNEMPCLMS     Average weekly initial claims for unemployment insurance   
              (Mn)                                                       

UNEMPRATE     Unemployment rate (percent of civilian labor force)  New   
              series 1/94                                                

Note: the retail sales series (whose names all begin with RS) are the series which are to be fitted with forecasting models; the other variables are included as potential independent variables. You may also obtain other series from Datadisk if you wish.