Methodology

We defined the business objective as reducing stockouts while lowering emergency replenishment costs. Forecasts were scoped at the warehouse × SKU level with a 7-day planning horizon and weekly update cadence.

The evaluation metric was MAPE, with a secondary focus on bias and gross inventory error.

Data was merged from sales history, inbound shipments, promotions, and event calendars. Missing values were imputed using forward fill for demand and zeros for promo flags. External signals such as weather index and public holidays were added to capture demand drivers.

A Prophet baseline model captured seasonality and trend. Residuals were modelled with LightGBM using product, location, and event features. The final forecast was generated by blending the two outputs to capture both structural and promotional demand shifts.

Forecast bias was measured against holdout windows and corrected using a scaled bias adjustment. For SKU categories with intermittent demand, an intermittent demand correction layer reduced false positive inventory. A safety stock margin was added based on lead-time uncertainty.

Walk-forward cross-validation was used across 12 rolling weekly windows. Forecast accuracy was evaluated on both aggregate SKU demand and top 20 high-volume items. The pipeline also tracked prediction intervals and the proportion of forecasts within tolerance bounds.

Daily forecasts are generated and published to a Power BI dashboard for supply planners. Alerts are triggered for demand spikes, stockouts, and supplier lead-time deviations. Model drift is monitored using percentage error and inventory turnover signals.