A Multi-Task Prediction Framework for Sales Prediction

Sales prediction has become essential for making promotion decisions while various types of influential factors bring great challenges to predicting sales accurately. In the paper, we propose a multi-task sales prediction framework (MTSP) which consists of two major stages. Firstly, we apply different feature extraction structures to obtain feature representations of different types of influential factors. Then, we introduce a multi-task prediction architecture to guide the sales volume forecasting and the seller predicting task. Experiments are conducted on the dataset constructed from real-life gasoline station data and experimental results demonstrate that our proposed framework achieves the lowest prediction error among baseline methods.