Cluster-based Cross-Learning on Decomposed Components: A Novel Framework for SKU Sales Forecasting

Accurate sales forecasting at the SKU level is crucial for retailers to optimize inventory management and improve supply chain efficiency. However, the high dimensionality and sparse data of SKU sales pose significant challenges for traditional time series forecasting methods. To address these issues, this paper proposes a novel forecasting framework that integrates time series clustering, STL decomposition, cross-learning, and ensemble techniques in a rolling mechanism. The framework first employs Dynamic Time Warping (DTW) based K-Means clustering to group SKUs with similar sales patterns. Each SKU's sales are then decomposed into seasonal, trend, and remainder components using STL decomposition. Next, a cross-learning approach using the XGBoost algorithm is applied within each cluster and component to share information across related SKUs. Finally, the forecasts of different components are ensembled to generate the final prediction. Experiments on real retail sales data of 100 SKUs demonstrate that the proposed method significantly outperforms comparative methods, including non-clustering, non-decomposition, individual modeling, and the classical ARIMA approach, in terms of MAPE over a 4-week forecasting horizon. The results highlight the effectiveness of the proposed framework in improving retail sales forecasting accuracy by leveraging cross-series knowledge via clustering and decomposition techniques.