Method
The first declared periods form the initial training set. The tool forecasts the next period, reveals its actual demand, expands the training set, and repeats. It ranks historical mean, Croston, SBA, and TSB by mean absolute error; absolute mean bias breaks a tie.
Why this is a separate decision
A fitted forecast answers “what baseline does this method produce now?” A rolling backtest answers “did this method earn trust on unseen periods?” A low error on one sparse SKU can be luck, so repeat the test across comparable parts and operational horizons.
Limits
One-step MAE weights every unit equally and does not encode outage consequence, asymmetric stockout cost, lead-time aggregation, or inventory service. Fixed smoothing inputs are not optimized. Do not compare methods on censored sales history or tune parameters against the reported test periods.