inventory-demand-planning

Select the method based on demand pattern: weighted moving averages for stable items, Holt's double exponential smoothing for trending items, Holt-Winters for seasonal items, and Croston's method for intermittent demand with over 30% zero-demand periods. Use ABC/XYZ classification to match method rigor to item value.

Use the formula SS = Z × √(LT_avg × σ_d² + d_avg² × σ_LT²) to capture both demand and lead time variability. Vendors with lead time coefficient of variation above 0.3 can require safety stock 40–60% higher than demand-only formulas suggest.

Use WMAPE (weighted mean absolute percentage error) instead of MAPE for low-volume items, since MAPE inflates errors when actuals are near zero. Reserve MAPE for items averaging 50+ units per week, and track bias to detect systematic over- or under-forecasting.

Use analogous item profiling: select 3–5 similar items matched on price point, category, brand tier, and demographic, then borrow their demand variability as a proxy. Add a 20–30% safety stock buffer for the first 8 weeks and taper as the item accumulates its own history.

ML methods like gradient boosting are only justified with large data (1,000+ SKUs and 2+ years of weekly history), multiple external regressors, and an ML engineering team for monitoring. Without quarterly retraining, model drift erodes the 10–20% accuracy advantage over simpler methods.