Prediction Intervals

Introduction

Prediction intervals for regression tasks provide a range within which the actual outcome is expected to fall with a given probability. They add a layer of transparency and confidence to model predictions by indicating the uncertainty associated with these predictions.

Prediction intervals are defined in Analysis > Basic > Metrics > Uncertainty.

They can also be used as input for prediction overrides

How does it work

To generate prediction intervals, we employ a helper model alongside the primary predictive model. This helper model is trained to estimate the bounds of the prediction interval, by fitting the residuals (i.e. the differences between observed and predicted values) from the primary model while taking into account the coverage level that you specified.

Coverage level

The default coverage level is set at 95%, meaning that the prediction intervals generated is estimated to contain the actual outcome 95% of the time, under the model’s assumptions. You can adjust this level based on your requirements for intervals with a higher probability of containing the actual outcome (resulting in wider intervals) or with a lower probability of containing the actual outcome (resulting in narrower intervals).