Feature Selection Strategies for Machine Learning-Based Demand Forecasting in Joint Distribution Centers for SMEs*

Purpose: This study aims to develop a machine learning–based forecasting model tailored to the operational characteristics of regional joint distribution centers used primarily by small retail stores. Unlike large-scale fulfillment centers, these hubs face highly irregular and low-volume demand at the SKU level, making accurate forecasting particularly challenging. Furthermore, important predictors including promotions and store-level operational details are not accessible in this setting, creating structural constraints for effective variable selection. To address these challenges, this study proposes a forecasting framework that leverages demand information from related product groups as proxy variables and evaluates two feature selection strategies—Wrapper and Filter methods—to identify a scalable and accurate prediction approach. Research design, data, and methodology: Using real transaction-level sales data from a Korean regional distribution center covering January 2021 to December 2024, daily sales were aggregated into 209 subcategories. A Random Forest model served as the common predictive algorithm across three conditions: (1) a benchmark model using only self-lagged features, (2) a Wrapper approach employing backward stepwise selection based on feature importance, and (3) a correlation-based Filter approach selecting top-N subcategory features via Pearson correlation. Model performance was evaluated using RMSE on a hold-out validation set, and total computational time, which includes feature selection and model training process, was compared across methods. Results: The Filter (correlation-based) method significantly outperformed the Wrapper method in both computational efficiency and predictive accuracy. The Filter approach required 3 to 100 times less computation time than the Wrapper method and produced lower RMSE, with optimal performance observed when using approximately 300 top-correlated features. Performance gains were particularly pronounced for subcategories exhibiting low sales volume, intermittent demand, and strong seasonal or co-movement patterns. In contrast, the Wrapper method struggled with high-dimensional feature spaces and failed to identify optimal feature sets within reasonable time. Conclusions: The findings indicate that in joint distribution center environments, correlation-based Filter selection is highly suitable, particularly in logistics settings characterized by sparse, volatile, and interdependent product demand. This approach simultaneously delivers accuracy and computational efficiency—both essential for operational decision-making in regional distribution hubs. Overall, the study contributes a domain-specific, data-driven methodology that enables more reliable shipment forecasting, improves fulfillment efficiency, and provides an analytical foundation for developing digital fulfillment models to enhance the competitiveness of small and medium-sized retailers.