A Hybrid LSTM-Prophet Model for Sales Forecasting and Inventory Optimization in E-Commerce Time Series

Abstract

With the rapid development of e-commerce, accurately predicting product sales and achieving dynamic inventory control have become key challenges for enterprises to optimize supply chain management. The traditional Prophet model excels at capturing long-term trends and seasonal characteristics in time series, but has limited modeling capabilities for complex nonlinear relationships. Although LSTM neural networks can effectively learn the dynamic dependencies of sequential data, they tend to overlook explicit temporal patterns. This study constructs an LSTM-Prophet fusion model, first utilizing Prophet to decompose the trends, seasonality, and holiday effects of sales data, and then inputting the residuals into LSTM for nonlinear correction. Finally, the prediction accuracy is improved through weighted fusion. Experiments based on 150,000 daily sales data from an e-commerce platform over three years show that the average absolute error of the fusion model is reduced to 8.7, which is 29.3% and 17.1% lower than that of the single Prophet (12.3) and LSTM (10.5) models, respectively, and the root mean square error decreases by 22.6%. In inventory control simulations, this model drives an 18.4% increase in inventory turnover rate and reduces the out-of-stock rate to 3.2%, effectively balancing prediction accuracy and computational efficiency. However, the model still relies on high-quality historical data, has high computational complexity, and has limited adaptability to new product launches and emergencies.