Demand Forecasting Engine
Multi-horizon demand forecasts converted into inventory policy via quantile regression.
Challenge
Ordering average expected demand means ~50% stockout probability under skewed demand. Point forecasts ignore risk and produce fragile inventory.
Solution
Walk-forward validated XGBoost quantile regression (4.01% WAPE, 31.7% better than seasonal naïve) across 45 stores at 1, 2 and 4-week horizons, converted to policy by formula — safety stock = Q90−Q50, reorder point = expected lead-time demand + safety stock for ~90% coverage. A Streamlit decision dashboard per store with 80% prediction intervals and stress scenarios for growth, volatility and supply delay.
View GitHub Repo
