In this section, we will compare the ?ve forecasting methods using the case study data described in Section 4. Methods 1-3 will ?rst be compared for the full data set (assortment groups 1-3) and in their forecasting accuracy for Season 2 (based on Season 1) and Season 3 (based on Season 2). Then, for Assortment group 1 and Season 3, Methods 1-3 will also be compared to Methods 4 and 5 based on expert judgment.

We used three different performance measures of forecast accuracy: mean absolute percentage error (MAPE), mean absolute deviation (MAD) and mean percentage error (MPE). The comparative performance of the different methods was consistent over the three error measures, and hence we report the results for MAPE only. We remark that we did not consider the mean square error (MSE) because of its sensitivity to outliers.

We report overall results (per year) as well as for classes of SKUs that are based on the value of the preview demand P. The classes that we use are P =0, 0< P 6 2, 2 < P 6 5, 5 < P 6 10, P > 10. We use this classi?cation, since we expect that higher preview demands will imply more reliable statistical extrapolations of that demand, and will therefore in?uence the relative performance of the different methods, especially statistical (Methods 1-3) versus expert judgment (Methods 4-5). We remark that the overall results exclude SKUs with P = 0, i.e. report the average over all SKUs with positive preview demand. This is done because some of the methods will always result in a zero forecast for these SKUs, and their inclusion would therefore reduce the meaningfulness of the overall results.