Demand Forecasting

Demand forecasting predicts future customer demand for products and services, enabling optimized inventory management, production planning, workforce scheduling, and financial planning. This skill covers seasonal pattern detection, promotional lift modeling, new product forecasting, hierarchical forecasting across product/geography hierarchies, intermittent demand handling (Croston's method), and forecast reconciliation techniques.

## Key Points

1. Demand decomposition into base, trend, seasonal, and promotional components is the essential first step
2. Seasonal patterns should be detected empirically (autocorrelation, FFT) not assumed; the period may not be what you expect
3. Promotional lift should account for forward buying (post-promotion dip) to avoid overestimating incremental value
4. New product forecasting requires analogous products or diffusion models since there is no history to extrapolate
5. Croston's method (or SBA) is essential for intermittent demand; standard methods overpredict by ignoring zero-demand periods
6. Hierarchical reconciliation ensures that SKU forecasts sum to category forecasts; MinT optimal reconciliation minimizes total variance
7. Use Weighted MAPE (WMAPE) over MAPE for demand with zeros; track bias separately to detect systematic over/under-forecasting
8. The tracking signal is the practical early warning system: when it exceeds +/-4, the model needs retraining