Sales forecasting accuracy over time: An empirical investigatio

This study investigated forecasting accuracy over time. Several quantitative and qualitative forecasting models were tested and a number of combinational methods was investigated.;Six time series methods, one causal model, and one subjective technique were compared in this study. Six combinational forecasts were generated and compared to individual forecasts. A combining technique was developed.;Thirty data sets, obtained from a market leader in the cosmetics industry, were used to forecast sales. All series represent monthly sales from January 1985 to December 1989. Gross sales forecasts from January 1988 to June 1989 were generated by the company using econometric models. All data sets exhibited seasonality and trend.;Three accuracy measures were employed in the investigation. These are: (1) Mean percentage Error (MPE), (2) Mean Absolute percentage Error (MAPE), and (3) Root Mean Squared Error (RMSE).;Nonparametric statistical tests (Friedman nonparametric analysis of variance and the Wilcoxon matched-pairs signed-ranks test) were employed to test whether the difference in performance among the methods was due to chance. Given the relatively small sample, these tests are preferred to the parametric t-tests.;Results indicated that combining several quantitative techniques is better than individual methods; that a weighted average combining is better than a simple average combining; that quantitative methods are more accurate than subjective methods; that combining quantitative and qualitative methods provide better forecasts than the individual methods; that the level of accuracy depends on the time horizon of the forecast; and that causal models are superior to time series methods.;Future research should focus on the reasons for the differences in accuracy achieved by the different forecasting techniques. More quantitative and subjective methods should be investigated at both macro and micro levels. In addition, future research should focus on different demographic data to confirm the results of this study. Finally, the study suggests that better methods of averaging forecasts be investigated.