DataData
Analysis uses information that would not have been known
Typical symptoms
- Historical results look unusually smooth
- Signals depend on revised or future data
- Performance falls apart when timestamps are shifted
Likely causes
- Inputs were aligned to the wrong timestamp
- Corporate actions or classifications were applied retroactively
- Final-period values were used inside the period being tested
Checks
- Audit each input's availability time
- Shift features forward and backward to test sensitivity
- Review whether labels were known at decision time
Resolution path
- Use point-in-time datasets where possible
- Lag inputs conservatively
- Document any unavoidable timing approximation
Related terms: lookahead, timestamp, point-in-time, feature lag