Advanced explainability

To break down predictions into the drivers that shaped it:

1. After a forecast run, open the Advanced Explainability panel. 
2. Use the Choose an Item dropdown to select the item you'd like to explore. 
3. Use the tabs to switch between views and explore driver impact. 

Impact analysis ranks the drivers that most influence the prediction for your selected item, summarized across all time periods combined rather than a single point in time. For each driver: 

• Its effect is measured at each period.
• The direction (up or down) is set aside.
• The sizes are averaged into one number.

Each bar shows how big that driver's typical push is. A longer bar means it moves the prediction more. 

Impact Distribution shows how consistently a driver influences a prediction over time by plotting individual period contributions along a direction scale. Dots to the right push the prediction up, and dots to the left pull it down. 

Look for a clean color split from left to right. This means the driver has a consistent, dependable relationship with the prediction. 

Impact Dynamics plots the prediction for your chosen item over time. The prediction, displayed as a solid line, is overlaid with its strongest drivers, displayed as a dashed line. 

A reliable forecast has drivers that move with the prediction. Drivers that lead the prediction can act as an early warning, while drivers that lag behind the prediction tend to confirm a trend after the fact. When a driver and prediction suddenly diverge, this may be an anomaly or a structural change in the data that's worth investigating.

Single Prediction Impact breaks down how a single prediction was built for your chosen item. It starts at the baseline, then each driver pushes the running total up or down to reach that period's prediction. 

The drivers account for the baseline and each contribution, and match the prediction for the selected period. Move the time slider to change the breakdown from one period to the next.

The examples show how each view works in practice in macro rankings, direction consistency, and time periods. 

The Impact Analysis view displays the driver impact ranking for an air conditioner.  Six drivers are ranked by their average impact across all time periods. 

The linear trend is the biggest lever for this item. Its typical push is almost twice that of the next driver. The bottom two drivers still matter, but shifting them changes the prediction far less. To see which way and how consistently each of these drivers pushes, open Impact Distribution.

While Impact Analysis reveals scale, Impact Distribution reveals behavior. By plotting each time period as an individual point along a directional axis and coloring points by feature value, you can verify whether a driver's behavior matches physical domain logic. 

Forecaster has found three distinct driver behavior patterns for the air conditioner item. 

The Impact Dynamics view plots the final forecast line alongside key driver series over time. You can observe push-and-pull behaviors and flag anomalies where the forecast trajectory diverges from its underlying drivers. 

The air conditioner prediction curve climbs to a peak in April 2026, moderates through the summer months, and picks up again toward September. 

The Single Prediction Impact view breaks down a single point in time down to its mathematical components in January 2026. 

Five drivers push the prediction above the baseline, led by the linear trend and a related data driver, while one drive