Forecasting model selection for inventory cost reduction: insights on time-series features and model selection

Growing product variety has increased the burden of demand forecasting. This study proposes a method for selecting forecasting models suited to each product’s time-series characteristics and inventory-management properties, aiming to reduce inventory costs. This study constructed a light gradient boosting machine using time-series features of sales quantities, inventory management parameters, and forecasting-model labels as inputs, which outputs each model’s relative performance in terms of inventory cost compared to other models. The method was validated using a subset of the M5 competition data, together with various inventory management parameters, including lead time. To identify which time-series features drive each method’s relative superiority, Shapley additive explanation values were computed for each model. The proposed method achieved inventory cost and cycle service level performance comparable to that of the best-performing benchmark methods. Notably, simple exponential smoothing and the Kalman filter – which showed smaller variation in explanation values across features – performed stably regardless of time-series characteristics.