Article ID: | iaor20031791 |
Country: | United States |
Volume: | 31 |
Issue: | 4 |
Start Page Number: | 789 |
End Page Number: | 812 |
Publication Date: | Oct 2000 |
Journal: | Decision Sciences |
Authors: | Klein Gary, Jiang James J., Zhong Maosen |
Keywords: | neural networks, artificial intelligence |
Analyzing scanner data in brand management activities presents unique difficulties due to the vast quantity of the data. Time series methods that are able to handle the volume effectively often are inappropriate due to the violation of many statistical assumptions in the data characteristics. We examine scanner data sets for three brand categories and examine properties associated with many time series forecasting methods. Many violations are found with respect to linearity, normality, autocorrelation, and heteroscedasticity. With this in mind we compare the forecasting ability of neural networks that require no assumptions to two of the more robust time series techniques. Neural networks provide similar forecasts to Bayesian vector autoregression, and both outperform generalized autoregressive conditional heteroscedasticity models.