Residuals and prediction intervals
A series whose error grows with its level is split into training and test data. In-sample residuals, the errors of the model on its own training data, are small because the model has already seen those values: the interval built from them covers only 60% of the test observations for a nominal 80%. Out-of-sample residuals, obtained by backtesting a model trained on the first part of the training data over the rest of it (validation), are larger and the interval covers 80%. Binned residuals, grouped by predicted value, keep that coverage (80%) with an interval that is narrow for low predictions and wide for high ones.
Residuals and prediction intervals
1
Residuals
2
In-sample
3
Out-of-sample
4
Binned
skforecast.org
0.2 s