Sales forecasting in rapid market changes using a minimum description length neural network

In this study, we address a real sales forecasting problem of a multinational fashion retailer. The fashion retailer has been operating retail stores and warehouses in a major Asian city for decades. The volume of daily sales of each stock keeping unit in this city needs to be determined to develop an effective and efficient inventory and logistics system for the retailer. Based on complicated real sales data sets, we use a minimum description length neural network (MDL-NN) which searches for the optimal model size to predict this value. In order to address the problem of intermittent sales data and zero actual sales, we propose a revised mean absolute percentage error (RMAPE) measurement for performance evaluation. Our experimental results show that for most test data sets, the evaluation results based on RMAPE are consistent with those based on two other established measurements, the symmetric mean average percentage error and mean absolute scaled error. Specifically, on all three performance indicators, the MDL-NN method achieves a smaller error than almost all other commonly used sales forecasting methods on almost all test data sets. Finally, we examine the gap between the performance indicators between the forecasting result values in decimal format, which are directly generated by the forecasting methods, and the forecasting result values in integer format, which are rounded off from the decimal numbers. Our findings indicate that this gap is not trivial and should not be ignored.