Time series and forecasting · insight-1205
End each backtest origin on a complete aggregation period
- Evidence type
- practitioner heuristic
- Difficulty
- easy
- Roles
- Data scientist, Data analyst, ML engineer
- Industries
- retail, aviation, e-commerce
Summary: situation, action, reason, boundary
Situation
Daily bookings are rolled up to weeks, and the cutoff falls on a Wednesday; the training week's total contains three days and the test week begins mid-week.
Action
Align every cutoff with the end of the coarsest aggregation period used by features or targets (week, month, fiscal period). If the business runs on a different week boundary, use that one.
Reason
Aggregates are only defined on complete periods. A cutoff inside a period either truncates the last value or lets post-cutoff days enter a feature.
Boundary
Multi-granularity pipelines: daily data with weekly or monthly features, fiscal calendars, 4-4-5 retail months.
Symptom or risk
A partial last period looks like a demand drop to the model, and a week total that straddles the cutoff mixes known and unknown days.
Exceptions
Daily models with no coarser features need no alignment.
Trade-offs
Fewer possible origins; production forecasts issued mid-period need an explicit partial-period feature instead.
Pitfalls
Aligning the training window but not the label window, so the first test week is only partly scored.
Example
Hypothetical: a Wednesday origin with weekly features sees a half-week of 300 bookings against a usual 700; a naive model reads that as a 57% collapse.