Primacy of feature engineering over architectural complexity for intermittent demand forecasting.

Area of Science:

• Operations Research
• Supply Chain Management
• Statistical Modeling

Background:

• Intermittent demand forecasting presents significant challenges in large supply chains due to data sparsity and variability.
• Existing research often focuses on complex model architectures, overlooking statistically grounded feature engineering.

Purpose of the Study:

• To introduce the Smoothed Hybrid Occurrence-Size (SHOS) framework for enhanced intermittent demand forecasting.
• To evaluate the effectiveness of SHOS-generated features in supervised machine learning models.

Main Methods:

• Developed the SHOS framework using sparsity-aware exponential smoothing for demand occurrence and size estimation.
• Incorporated SHOS features into machine learning models trained on large-scale, zero-padded panel data.
• Validated the approach on an automotive aftermarket dataset using rolling-window cross-validation.

Main Results:

• SHOS-enhanced models reduced Mean Absolute Error (MAE) by ~50% and Weighted Mean Absolute Percentage Error (WMAPE) by >40% in intermittent demand segments.
• The single-stage SHOS framework outperformed complex two-stage hurdle-based models.
• Statistical testing confirmed the significant performance advantage of the SHOS model (p < 0.001).

Conclusions:

• Statistically informed feature engineering, as demonstrated by SHOS, can be more effective than increased model complexity for intermittent demand forecasting.
• The SHOS framework offers a computationally efficient and interpretable alternative for large-scale operational deployment.
• Future research should explore SHOS validation across diverse application domains.