An Aggregate Store Sales Forecasting Framework based on ConvLSTM

Aggregate sales forecasting means predicting the total sales over product units, locations or time buckets. It could help companies make production plans and marketing decisions more precisely. Various techniques have been used in sales forecasting, while they all treat the forecasting objects separately. However, there is spatial correlation between sales of neighboring stores or areas, for which existing methods lack of consideration. In this paper, we propose a novel model for aggregate sales forecasting based on the deep learn- ing technique ConvLSTM, which replaces the Hadamard production by convolution operation comparing to traditional recurrent neural networks. The ASFC framework could take spatial correlations between neighboring stores into account when conducting sales forecasting. To verify the effectiveness of our proposed framework, we use real world retailing sales data to implement baseline ex- periments and compare the model with other benchmark methods. The experimental results show that the proposed model effectively improves the prediction performance, which provides a new path of aggregate sales forecasting for both industry and academia.