ML-Based Spare Parts Demand Forecasting
Developing machine learning models to forecast intermittent demand for semiconductor equipment spare parts.
Spare parts for semiconductor equipment exhibit classic intermittent demand patterns — long periods of zero demand punctuated by sporadic, high-value orders. Traditional forecasting methods like moving averages fail badly with this demand profile, leading to either chronic stockouts or excessive inventory.
Comparing multiple approaches: Croston's method (the traditional intermittent demand baseline), gradient-boosted trees using part characteristics and installed base data, and ensemble methods that combine statistical and ML predictions.
Building warehouse capacity forecasts using S&OP data and long-term growth trends.
Work in progress — early results show 30%+ improvement in forecast accuracy over traditional methods for key part categories. The forecasting models integrate with inventory optimization to automatically recommend safety stock levels.
“For intermittent demand, the installed base is your best predictor. Knowing which machines are in the field, their age, and their utilization patterns tells you more than years of demand history.”