Leakage-safe lagged features

How would you construct a lagged rolling mean feature for a next-day return model without accidentally using the predicted day's return?

Define the decision timestamp first. Use only returns observable by that timestamp, group by asset, and apply any required shift before rolling. Sort within assets and prevent windows crossing entity boundaries. Verify a small example manually around the split cutoff. A rolling function's default alignment may include the current row, which is valid only if that row's value is already observable at prediction time.