A LSTM Approach for Sales Forecasting of Goods with Short-Term Demands in E-Commerce

Abstract

This study proposed a model to forecast short-term goods demand in E-commerce context. The model integrated LSTM approach with sentiment analysis of consumers’ comments. In the training stage, the sales figures and comments crawled from “taobao.com” were preprocessed, and the sentiment rating of comments were analyzed for “positive”, “negative” and confidence. The LSTM model was trained to learn the prediction of future value according to the time-series sequence of sales and sentiment rating of comments. Due to the characteristics of short-term goods, there are not enough history data to evaluate cyclic and periodic variation, so the decision makers have to react to market conditions and take appropriate actions as soon as possible. It also suggested that to adjust the weight of sentiment rating appropriately could further improve the forecasting accuracy. The study fulfilled the goal for supporting them to make use of minimal trading data to achieve maximal predictive accuracy. The results demonstrated that the proposed LSTM approach performed high-level accuracy for sales forecasting of goods with short-term demands.