Research on a Lightweight Supply Chain Demand Forecasting Model Based on STL Decomposition and Dynamic Feature Selection

Abstract

The accuracy of supply chain demand forecasting directly impacts inventory cost control and operational efficiency optimization.
Traditional time series models struggle to handle the coupled effects of complex seasonal fluctuations and unexpected events, while existing
machine learning methods often suffer from overfitting due to noise introduced by redundant features. To address this challenge, this study
proposes a lightweight ensemble forecasting framework integrating STL time series decomposition with SHAP-based dynamic feature selec
tion. The method first employs STL decomposition to uncouple sales data into trend, seasonal, and residual components. Subsequently, it dy
namically selects high-contribution features through a rolling-window feature evaluation mechanism based on SHAP values, effectively miti
gating noise impacts. Experiments validated using the Rossmann pharmaceutical sales dataset demonstrate that the proposed method achieves
a WMAPE of 6.2%, representing a 3.1% reduction compared to a standalone XGBoost model, while improving the R metric to 0.956. Ad
ditionally, the framework exhibits 12.6 times higher training efficiency than LSTM models, showcasing superior computational performance.
Ablation studies further confirmed that the trend slope feature exhibits the highest predictive contribution, with an average SHAP value of 0.42.
The proposed framework combines high computational efficiency with strong interpretability, making it particularly suitable for resource
constrained applications in small and medium-sized enterprises. This method not only effectively improves supply chain demand forecasting
accuracy but also extends its methodology to sectors with similar demand characteristics, such as retail and fast-moving consumer goods.