A cost-aware intermittent demand forecasting method combining temporally attenuated discrete gray model with reward-guided adjustment

Purpose Intermittent demand in distribution-center operations is difficult to forecast because long zero-demand periods are often interrupted by irregular demand bursts. This study develops TADGM–RGPL, a cost-aware gray forecasting framework that combines a temporally attenuated discrete gray model (TADGM) with reward-guided parameter learning (RGPL), to account for both statistical calibration deviation and the asymmetric inventory consequences of underforecasting and overforecasting. Design/methodology/approach TADGM introduces a bounded recency-sensitive driver directly into the DGM state-transition equation, while RGPL learns a scalar decision-calibration factor from pre-origin inventory-cost feedback. The framework is evaluated on 701 intermittent and lumpy DC-SKU series using 4,206 rolling-origin windows. Conventional intermittent-demand methods, gray models, XGBoost and LightGBM are included as benchmarks. Findings TADGM does not dominate the benchmark methods in point-forecast accuracy. Standard TADGM–RGPL significantly reduces inventory cost relative to the conventional intermittent-demand and gray benchmarks under the stockout-sensitive p/h = 10 setting, whereas XGBoost and LightGBM achieve lower overall cost when all series are pooled. ADI-stratified analysis shows that TADGM–RGPL attains the lowest average cost in the highest-intermittency quartile, although its advantage over the machine-learning benchmarks is not statistically significant within that subset. Originality/value The study differs from residual reweighting by introducing recency sensitivity as a time-varying gray input in the DGM state-transition equation and by separating statistical forecasting from downstream cost-aware decision calibration. The empirical analysis further identifies the demand conditions under which the proposed forecasting-decision framework is relatively more competitive.