A multiple long short-term model for product sales forecasting based on stage future vision with prior knowledge

Daifeng Li

Abstract: Deep neural network (DNN) based multivariate time series (MTS) forecasting has been widely studied in many domains. The approach has also been successfully applied to product sales forecasting, which is invaluable in the strategic development of enterprises. However, one key challenge is to efficiently combine all influential factors into a unified framework by considering their long short-term correlations. This is particularly challenging for real time sales predictions with many important features unknown from the perspective of future vision, because of the complex and dynamic changing environment of sales time series. Besides, DNN based methods could not effectively capture stable linear correlations between multivariate sales time series. To address these challenges, a novel Stage future-vision-based multiple Long Short-term model with prior knowledge (SLST-PKNet) is proposed. The model constructs sub-models according to the types of different influential factors, and a stage future vision mechanism is used to model dependent correlations between future influential factors and product sales by combining a two-layer convolutional neural network (TLCNN) and a two-stage LSTM (TSLSTM) into a unified framework. The combination of TLCNN and TSLSTM is the basic component for long short-term modeling, and is an effective solution for capturing dynamic patterns by fusing the outputs from different layers and stages. A dynamic co-integration mechanism (DCI) is introduced to capture strong correlations between time series, which DNN is not good at. In order to further improve the capability of the model to capture long short-term patterns from complex environment, domain prior knowledge (PK) is integrated as supervision information. Extensive experiments are conducted on two sales datasets: Galanz and Cainiao. SLST-PKNet achieves significant performance improvements over 11 state-of-the-art baselines. The proposed model is also evaluated on two new datasets: Traffic and Exchange-Rate, to further verify its generalization capability.