Explainability

Explainability features help generate more accurate forecasts. Not all features impact forecast results in the same way across different datasets, time periods, and algorithm configurations.

Two types of features can be generated from an input data collection:

• Historical data: For the historical data in your data collection, Forecaster algorithms automatically generate underlying features based on seasonality and trend. If you select the Forecaster built-in holiday calendar, features are automatically generated based on the calendar.
• Related data: If you provide related data as part of the data collection, the features themselves and the lags of those features over time may be selected by the algorithm. Some examples include promotions and inventory.

If Forecaster determines a feature contributes significantly to the performance of a given algorithm, it's added automatically to the Forecast Feature Names list in the explainability results module.

Standard explainability helps you understand how specific features affect overall forecast results. It's available in the standard tier.  

Standard explainability helps you understand the underlying drivers of a Forecast model, captured in the explainability results module. For each item, it shows the relative impact of a given driver in relation to other drivers. 

Take a Forecaster data collection with information on historical oil prices, e-bike prices, and promotion data. With explainability, you can show whether ‌increased oil prices or promotions have a relatively greater impact on sales.

Available in the advanced tier, advanced explainability breaks down each prediction into the specific drivers that shaped it. It shows which drivers pushed a prediction up or down, by how much, and when. It produces per-period driver contributions in forecast units that add up to the predicted value for each item and period. 

Advanced explainability is available on the advanced tier. 

The contributions are produced in forecast units and equal the sum of the predicted value for each item and period.

Contributions describe how the model built a prediction in relation to its baseline. They show the direction and relative magnitude of each driver's influence for the selected item.