E-commerce Sales Prediction and Analysis Based on Multi-layer LSTM and Interactive Visualization Techniques

This study proposes a multi-layer long short-term memory (LSTM) model integrated with an attention mechanism for e-commerce sales forecasting. The experimental results show that the proposed method achieves mean absolute error (MAE) [Formula: see text] 12.45 and mean absolute percentage error (MAPE) [Formula: see text] 3.21% on the test set. The proposed method is superior to traditional methods such as single-layer LSTM and autoregressive integrated moving average (ARIMA). Based on the task experiment of 40 users (20 business personnel and 20 technical experts), in comparison with the baseline scheme (e.g., traditional static visualization and manual analysis process), the system’s average task completion efficiency has increased by 45.3%. Meanwhile, the satisfaction score of the visualization module reaches 4.8 points (out of 5), significantly improving the user’s understanding and operation experience. This study provides a cost-effective predictive and analytical solution for small e-commerce platforms (platforms with limited computing resources and fewer than 100,000 stock keeping units (SKUs)). Moreover, the study enhances prediction accuracy and interpretability through an interactive visualization tool.