Retail has two costly mistakes: empty shelves and excess stock piling up in the warehouse. Demand forecasting is how you balance the two based on data. So where should you start?
Start with the right data
A good forecasting model starts with well-prepared data. The key inputs are:
- Sales history by product, store and day
- Price changes and the campaign calendar
- Calendar effects such as public holidays, religious holidays and school terms
- Stock-out records; otherwise demand that could not be sold looks like “there was no demand”
Move from simple to complex
Every project should start with a simple baseline model, such as the same week last year or a moving average. Then statistical methods such as exponential smoothing (ETS) and ARIMA are tried, and where there are thousands of product-store series, machine learning models such as gradient boosting. A complex model is only valuable if it meaningfully beats the baseline.
Validate in time order
Forecasting models should be tested as if they did not know the future: the model is trained on data up to a point in the past and tested on the following period. Because percentage error (MAPE) can be misleading for low-volume products, weighted error measures (such as WAPE) give more reliable results.
Connect the model to operations
A forecast creates no value unless it is reflected in ordering and inventory decisions. The model needs to run automatically on a schedule, its results need to feed into ordering processes, and forecast accuracy needs to be monitored with a Power BI report.
Conclusion
Demand forecasting is not a one-off model but a continuously improving process. Starting with a small product group and measuring the results is the safest path before expanding the scope.