A two stage statistical framework for cold start spare part demand forecasting

Accurate demand forecasting for spare parts under true cold-start conditions remains a fundamental challenge due to extreme demand sparsity, zero inflation, and the complete absence of historical demand information. Conventional time-series methods, single-stage machine learning models, and sequence-based probabilistic forecasters are inherently ill-suited to this setting, as they either rely on historical observations or fail to properly represent zero-demand events and distributional uncertainty. To address this gap, this study proposes a novel Zero-Inflated Gamma Monte Carlo (ZIG MC) framework that explicitly decomposes demand into occurrence and magnitude components and generates fully probabilistic forecasts suitable for risk-aware inventory decision-making. The proposed approach integrates a Bernoulli classifier for demand occurrence with a Gamma-based magnitude model and employs Monte Carlo simulation to construct predictive demand distributions. Model performance is evaluated using a strict part-level nested cold-start validation protocol on an industrial transactional dataset, ensuring genuine generalization to previously unseen parts. Results demonstrate that ZIG MC consistently outperforms strong single-stage regressors, statistical hurdle models, and a state-of-the-art probabilistic deep learning benchmark (DeepAR). The proposed framework achieves the lowest point forecast error (MAE = 5.65), representing a 6.4% improvement over the strongest single-stage baseline, while also delivering superior scale-independent accuracy (MASE = 0.87). Probabilistic evaluation using the Continuous Ranked Probability Score shows an order-of-magnitude improvement over DeepAR (CRPS = 3.27 vs. 15.41), indicating substantially better calibration and sharper predictive distributions under cold-start conditions. Quantile-based reliability analysis confirms that predicted service-level quantiles are well aligned with empirical outcomes, enabling reliable translation of forecasts into inventory policies. Statistical significance testing further confirms that the observed performance gains are robust and not attributable to random variation. Sensitivity analyses demonstrate that forecasting performance is stable with respect to distributional assumptions and Monte Carlo sampling size. Inventory simulations reveal that ZIG MC yields higher fill rates and lower stock-out risk than competing probabilistic models at comparable inventory levels, directly linking improved probabilistic calibration to operational benefits. Collectively, these findings establish ZIG MC as a robust and practical framework for cold-start forecasting of intermittent demand, offering a principled foundation for uncertainty-aware inventory planning in data-scarce environments.