Prediction Errors
Prediction Errors
We assume the train and test sets have a balanced sample of difficult-to-predict-examples for a given model and performance metric.
Is there evidence that examples in the train set and test set are equally challenging to predict?
- Are there examples in the train set or the test set that are always predicted incorrectly?
- Are there domain segments of the train set or the test set that the are harder to predict than other segments?
- Are there some classes that are more challenging to predict than others on the train or test sets (classification only)?
- Are there deciles of the target variable that are more challenging to predict than others on the train or test sets (regression only)?