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

Area of Science:

• Operations Research
• Statistical Modeling
• Supply Chain Management

Background:

• Accurate spare parts demand forecasting under true cold-start conditions is challenging due to data sparsity and zero-inflation.
• Conventional methods fail to address the lack of historical data and distributional uncertainty inherent in cold-start scenarios.

Purpose of the Study:

• To propose a novel Zero-Inflated Gamma Monte Carlo (ZIG MC) framework for probabilistic demand forecasting in cold-start environments.
• To enable risk-aware inventory decision-making by generating reliable predictive demand distributions.

Main Methods:

• The ZIG MC framework decomposes demand into occurrence (Bernoulli classifier) and magnitude (Gamma model) components.
• Monte Carlo simulation is employed to construct predictive demand distributions.
• A strict part-level nested cold-start validation protocol was used on industrial transactional data.

Main Results:

• ZIG MC outperformed single-stage regressors, hurdle models, and DeepAR, achieving the lowest point forecast error (MAE=5.65) and superior scale-independent accuracy (MASE=0.87).
• Probabilistic evaluation showed an order-of-magnitude improvement in Continuous Ranked Probability Score (CRPS=3.27 vs. 15.41) over DeepAR.
• Inventory simulations demonstrated higher fill rates and lower stock-out risk with ZIG MC compared to other probabilistic models.

Conclusions:

• The ZIG MC framework provides a robust and practical solution for cold-start forecasting of intermittent demand.
• It offers a principled foundation for uncertainty-aware inventory planning in data-scarce environments.
• The model's performance gains are statistically significant, robust to distributional assumptions, and translate to tangible operational benefits.