Glossary
Data leakage
Information from outside the training period or the true information set that finds its way into a model's inputs or its testing, inflating results.
Leakage means a model sees something at training or test time that would not have been known when a decision was made. Common forms: features built from future data, scaling with statistics computed on the whole sample, and test samples whose labels overlap in time with training samples.
Leaked results look excellent and then fail in live use.
Example: with daily samples labelled by the return over the next 10 days, the label of one day shares 9 of its 10 days with the next day's label, so a random split puts almost the same outcome in both training and test sets.