From passive audit to active control : operationalising explainable AI in supply chain management for intermittent & lumpy demand : can SHAP-based explanations be translated into adaptive, rule-driven safety stock policies that optimize the cost-service trade-off for intermittent and lumpy retail demand?

Abstract

Retail inventory management faces a persistent tension between forecast accuracy and operational decision-making. Machine learning models such as gradient-boosted trees have demonstrated superior accuracy over traditional statistical methods on large retail datasets, yet inventory costs frequently fail to improve when these models are paired with static stocking rules that discard the contextual information the forecasts contain. This thesis investigates whether that gap can be closed by extracting and operationalising the explanatory content of a machine learning forecast, rather than treating it as a black-box input to a fixed policy.

M5 Walmart dataset is used in this study, including California state, Hobbies department, 2162 SKUs having only intermittent + lumpy demand. Further we develop & evaluate a SHAP-based adaptive inventory policy. Machine learning model XGBoost is trained and then decomposed model’s predictions using SHAP (SHapley Additive exPlanations), attributing each forecast to the individual features driving it. Calendar and event features, including holidays, SNAP benefit days, and promotions, produce large, precisely timed contributions on the days they are active despite modest global importance.

To optimize replenishment for intermittent product portfolios, we aggregate local feature attributions into a daily, SKU-level risk score that drives a bidirectional safety stock policy. This system increases inventory buffers during the highest-risk 3% of days and strategically trims stock across the remaining 97%. This adaptive approach cut total inventory costs by 8.0% as compared to a static machine-learning baseline and common calendar knowledge. it secured a 92.2% demand fill rate, while exceeding the 90% operational target. This financial improvement is statistically significant (binomial sign test, p < 0.0001) and holds true for 85.0% of individual SKUs. A baseline calendar heuristic applying identical stock boosts to all event days inflated costs by 10.3%, proving that granular SKU-level discrimination via SHAP rather than raw calendar awareness, drives the optimization. These savings remain robust across backorder-to-holding cost ratios of 3, 5, and 10, and hold directionally consistent across lead times from one to seven days. Ultimately, these results demonstrate that SHAP can function effectively as a live, automated inventory control signal rather than a backward-looking diagnostic, translating minor forecasting improvements into systemic, broad-based cost reductions.