skip to content
Avoid leaking features
Avoid patterns that can incorporate a single leaking attribute many times (Aggregate frame, Aggregate subgroup, Ratio).
Always use nonzero blackout on temporal data (even just 1 second is enough).
Beware of tables that do not use time constraint. Potentially all predictors calculated from these tables are leaking!
Cross-validate over time (before passing the data into a model, split the data to training and testing part based on time).