Historical volume and POS data

The machine learning forecasting model relies heavily on historical sales or volume data to generate forecasting models. There should be at least 3 years (36 months) of historical data available for each driver that is using machine learning to generate forecasts. This data should be continuous during the period that forecasting models are created for, which is selected when configuring the Batch Forecast Training task. Stretches of missing data, for example entire weeks or months missing, reduce the accuracy of forecasts. For the best possible results, it is recommended to import 5 years (60 months) of historical data for each driver.

Sales and volume data are imported into WFM using the POS interface. In addition to importing historical data, this interface should be configured to regularly import new data as it is captured by the store POS systems.

See POS interface.

Machine learning forecasts can still be generated for drivers that have no historical data. This may be useful when new locations start up since the data required to derive a forecasting model is not yet available. The Batch Forecast Training task does not create a model for drivers where there is no related historical data in the database. A warning message is written to the job scheduler logs for drivers that do not have data. The forecast generated for these drivers is zero for all intervals, but it can be manually edited after forecast generation.