Most companies still forecast the way they did twenty years ago: a monthly number per product family, adjusted by gut feel, then split across regions using last year's ratios. It works — until a festival season, a new competitor launch or a fuel-price change hits one region harder than another.
Why granularity matters
Stock-outs don't happen at the national level. They happen at a specific dealer, for a specific model and fuel type, in a specific week. A forecast that can't see that level of detail can't prevent it. Forecasting by product × dealer × region × day sounds daunting, but modern machine learning handles it well — provided the data is right.
The five ingredients
- Clean, connected data — sales and dispatch history from ERP, dealer stock, schemes and promotions, holidays and local events. Real-time ERP sync matters: a forecast built on last month's data is already stale.
- The right models — gradient-boosted trees (XGBoost, CatBoost) are excellent at learning from many drivers at once; sequence models such as LSTM capture seasonality and trends. In practice, an ensemble usually beats any single model.
- Hierarchical consistency — dealer forecasts must add up to regional and national numbers so that production, sales and finance plan from one version of the truth.
- Scenario simulation — planners need to ask “what if we run this scheme in the south?” and see the impact before committing.
- A learning loop — every day's actuals feed back into the models, so accuracy improves over time instead of drifting.
Measuring accuracy honestly
“95% accurate” means nothing without a definition. Agree upfront on the metric (for example, 1 − weighted absolute percentage error), the level (SKU-dealer-week, or aggregated), and the horizon (next week vs. next quarter). Track it on a dashboard everyone can see. Honest measurement builds the trust that gets forecasts actually used.
Making forecasts useful
A forecast nobody acts on is just a number. The real value comes when forecasts drive decisions automatically: replenishment suggestions for each dealer, production plans aligned to regional demand, and early warnings when actual sales drift from the plan. Delivered through clear dashboards (we often use Power BI), forecasts become part of the weekly rhythm rather than a monthly report.
Getting started
Begin with one product line and two or three regions. Prove the accuracy uplift against your current method over a few months, then scale. The data work you do in the pilot — cleaning history, connecting ERP — pays off for every product you add afterwards.