Training the forecasting model

The machine learning forecasting model performs a training process to generate forecasting models for a location. The training process builds forecasting models by analyzing the data available for a location. This includes the historical sales or volume data and any additional data available for the location, for example historical promotion dates. When creating a forecast for the location, these forecasting models are used to predict volumes of each driver.

The training process is performed by the Batch Forecast Training job scheduler task. This task must be completed at least once for a location before you can generate forecasts for that location. When the task runs, it creates a forecasting model for each driver that is configured to use machine learning. This forecasting model is built based on the data available at the time of training. When configuring the task, you select the number of years of historical data that is used when generating forecasting models. In general, it is recommended to configure the task to use between three and five years of data. Using a greater number of years produces more accurate forecast results, assuming that continuous historical data is available for the selected period.

The Batch Forecast Training task should be set to run on a schedule. This allows the forecasting models to be recreated on an ongoing basis with new results data captured from the location's point-of-sale (POS) system. When the task runs for a location that has already been trained, new forecasting models are created for each driver that replace the previous models. For the most accurate forecast results, it is recommended to retrain the model with the same frequency that is being used to generate forecasts for a location. The training task should be scheduled to run before forecasts are generated to ensure up-to-date models are used to generate the new forecasts. For example, if a store generates weekly forecasts every Sunday night, the store should be retrained weekly on Saturday nights. Customers may prefer to retrain the model less frequently for performance reasons or to satisfy other requirements.

The Batch Forecast Training task specifies a location to be trained. Sub-locations of that location are also trained when the task runs. Sub-locations can include departments and drivers of a single store, or multiple stores that are sub-locations of a larger corporate entity. Due to the volume of data that must be analyzed, it may take some time for the task to generate each forecasting model. In environments with hundreds or thousands of stores, it is a good idea to create multiple tasks to divide stores into groups that you want to train together. Each task can be set to start on different days to ensure that training can be completed during an acceptable time window.