We implemented a forecasting engine that generates SKU-level demand and safety stock recommendations across channels and locations. The system blends classical time-series with transformer-based models to handle seasonality, promotions, stockouts, and new-item cold starts.
Planners get weekly updated forecasts, exception alerts, and suggested order quantities instead of wrestling with static spreadsheets. Inventory, operations, and finance all work from the same numbers, reducing stockouts while freeing up working capital.
The engine shifted planning from reactive fire-fighting to exception-based management, with planners spending time where forecasts and demand diverge the most.
Before the forecasting engine:
We introduced a modern forecasting stack:
The Demand Forecasting engine delivered:
The forecasting engine runs as a scheduled pipeline that pulls history, applies models, and writes forecasts and stock recommendations back to planning tools and dashboards.
Short-term (week) forecasts for operations, medium-term (month/quarter) for purchasing and budgeting.
Promotion flags, price changes, and campaigns are included as regressors to avoid over-forecasting after one-off lifts.
New items inherit parameters from similar SKUs and categories until they accumulate their own demand history.
Forecasts come with confidence intervals and override capabilities, keeping humans in control while using the model as a strong baseline.
The setup lets teams iterate on models and parameters without disrupting downstream planning processes or requiring planners to change tools overnight.
