Stock-Out Prediction for FMCG Products Using CatBoost: An Empirical Comparison of Class-Imbalance Strategies with Native Categorical Handling and Probability Ensembling

Abstract

Stock-out is a critical problem in fast-moving consumer goods (FMCG) supply chains because it directly causes lost sales and reduced customer satisfaction, yet its prediction is hindered by severe class imbalance. This study presents an empirical comparison of three class-imbalance strategies for CatBoost-based stock-out prediction: class weighting, leak-free SMOTENC applied within cross-validation folds, and no balancing. CatBoost is adopted for its native handling of high-cardinality categorical features, which avoids the dimensionality explosion of one-hot encoding. All strategies are tuned under an shared pruned hyperparameter-optimization protocol and evaluated on an FMCG dataset of 1.1 million transactions with an imbalance ratio of approximately 32:1. At F2-optimized thresholds, the three strategies perform within a narrow band (F2 of 0.8086, 0.7870, and 0.8075, respectively); a lightweight probability ensemble of the two strongest single models attains the best result (F2 = 0.8112, recall = 0.9058, PR-AUC = 0.7744), a modest gain of about 0.3% in F2 over the best single model. Notably, leak-free SMOTENC yields the lowest performance at roughly seven times the computational cost. The principal finding is therefore empirical rather than methodological: under a high imbalance ratio with a large absolute number of minority instances, synthetic oversampling offers no advantage over simpler strategies and incurs a substantial computational penalty. The main scientific contribution is a leak-free empirical benchmark that isolates the effect of the imbalance treatment and yields evidence-based guidance for large-scale, high-cardinality FMCG stock-out prediction; in practical terms, it shows that practitioners can rely on class weighting or a lightweight probability ensemble to achieve competitive detection while avoiding the roughly sevenfold computational overhead of synthetic oversampling.