Table gives the average MAPE for all SKUs with positive preview demand together (overall) and also per preview demand class. Furthermore, the error percentages in bold were signi?cantly lower (based on Tukey tests at a 5% signi?cance level) than those for other methods (if any) that are not in bold, but not signi?cantly different from each other.
It appears that all methods perform considerably better in Season 2 than in Season 3. An important contributing factor to the poor performance in Season 3 is that demand dropped sharply compared to previous years, although preview demand was comparable to previous years. This may have been caused by a number of factors, including macro-economic and weather conditions. We discussed this with company experts, but neither they nor we could identify important explanatory market or economic conditions as part of the cause.We remark that all data was collected before the start of the current global recession.
As expected, Method based on equal division performs worst on average. Method 1 (preview) provides the best overall performance. However, as is especially evident for Season 3, Method 1 can lead to large forecast errors for SKUs with high preview demand. Method 1 often results in much too large forecasts for those SKUs. This leads to large stocks remaining at the end of the season that either become obsolete or have to be sold below the cost price. Methods 3 avoids those large forecast errors for SKUs with high preview demand. Apparently, although high preview demand is indeed a reliable indicator of whether an SKU will be top, the exact ranking of the top SKUs based on preview demand is no guarantee that the ?nal ranking based on realized demand will be the same. This is illustrated for a speci?c product group with 9 SKUs in Table.
For this group, the three SKUs with the highest (lowest) preview demand indeed turn out to be the top (?op) SKUs. However, the realized demand for the SKU with the highest preview demand of 8, is only about half of that for the SKU with preview demand 7. For the above example with 9 SKUs, the three top SKUs and the three ?op SKUs are all correctly identi?ed. In general, especially for larger numbers of SKUs, the classi?cation is not perfect. However, most SKUs do typically end up in the correct class. To illustrate this, we consider a second example of a product group with 37 SKUs in Season 3. Table 4 shows the preview demand and the Forecast errors (MAPE) averaged over the SKUs of all three assortment groups for Seasons 2 (top) and 3 (bottom). All forecast methods are applied at the product group level. Errors in bold are signi?cantly lower (based on Tukey tests at a 5% signi?cance level) than those (if any) not in bold, but not signi?cantly different from each other.