Time series and forecasting · insight-1245
Give a boosted tree the same-weekday and same-season lags explicitly
- Evidence type
- practitioner heuristic
- Difficulty
- easy
- Roles
- Data scientist, ML engineer
- Industries
- retail, logistics, aviation
Summary: situation, action, reason, boundary
Situation
A daily boosted model uses lags 1 to 7 and a rolling 7-day mean, and its Monday forecasts keep missing the Monday level after a holiday week.
Action
Add lags at the seasonal period and its multiples, such as 7, 14, 21 and 28 days, same-weekday rolling means, and a calendar block. For annual comparison use 364 days to keep the weekday, subject to the horizon rule for lags.
Reason
A seasonal lag places the observation at the same position in the cycle directly into the feature set, so the tree can use it without having to infer the cycle from a time index.
Boundary
Daily and weekly boosted-tree forecasters with a stable seasonal cycle.
Symptom or risk
Trees see only the features they are given: after a holiday week the lag-7 value is itself distorted, and a trailing seven-day average blurs the weekday pattern, so no clean same-weekday reference is in the feature set.
Exceptions
Short histories cannot support annual lags; use calendar features and a seasonal-adjusted target instead.
Trade-offs
Longer lags remove the first rows of each series from training and add correlated columns.
Pitfalls
Using lag 365 for an annual comparison, which moves the weekday by one or two each year.