Product-level sales prediction in e-commerce using LSTM modeling

Sales forecasting remains an essential tool in market decision-makers’ arsenal. Reliable sales forecasting is crucial as it prevents stock-outs and excess inventories, but reliability in an e-commerce environment proves to be a challenge. This study proposes a Long-Short Term Memory (LSTM)-based model that forecasts at a product-level and generates a prediction interval by outputting lower and upper bounds, rather than a single-point estimate. The model utilizes a dual-input method, having a sequence of numerical features and a categorical product identifier as its input, to capture product characteristics. To ensure that the model generates a meaningful prediction interval, a custom loss function was designed focusing on the validity, coverage, and compactness of the interval. Using real-world Shopee e-commerce seller data, the model achieved a 92% coverage rate along with a moderate average prediction interval width, demonstrating a balance of accuracy and risk awareness. This approach provides a solution for scalable storewide prediction and risk-aware insights for better informed decision-making.