Time series and forecasting · insight-0241

Shift before rolling so window features exclude the target period

Situation
A 7-day rolling mean of sales is computed and joined to the same row as the target.
Action
Group by series, shift by at least the horizon, then compute trailing windows; assert in a unit test that each feature uses only timestamps before the origin.
Reason
Pandas rolling windows end at the current row by default, and centred windows include later rows, so a shift is needed to respect the origin.
Boundary
All lag and rolling features in pandas, SQL window functions and feature stores.

Symptom or risk

The window includes the target day itself, or centred windows include future days, so the model learns from the answer.

Future-known variables such as scheduled capacity may be aligned to the target row.

Larger shifts reduce recency for short horizons; build features per horizon if needed.

Rolling across series boundaries when the frame is not grouped by series.

Pandas rolling windows are trailing by default and can be centred, which places later observations in the window.
Windowing operations (pandas User Guide)