Questions / Quant research, Data & feature research
Keep feature and label horizons apart
A row at time t is labeled using returns over [t,t+5]. A training label’s outcome window extends into the validation period. What should be examined?
Answer it in the app It's free. You get hints, a second try if you miss, and a follow-up question.
Show the answer and worked solution
Answer: D. Check overlap between training label windows and validation information.
- A row labeled at t uses outcomes through t+5, so its information footprint extends beyond its feature timestamp.
- Compare those outcome windows with the validation interval. Purge conflicting training observations when the evaluation design requires independent information sets.
Interview shortcut. Label windows can leak across the split; purge the overlaps.
Takeaway. Split on the information footprint of labels as well as features.
Why the other choices fail
- Nothing; ordered feature timestamps rule out leakage. Label windows can leak even with ordered features.
- Check whether chronological sorting removes the overlapping return observations. Sorting changes row order, not which future returns were used to construct each training label.
- Check only that training feature timestamps precede validation feature timestamps. A training target can depend on later returns even when its feature timestamp is earlier. Feature ordering alone misses label overlap.
In the app this has a follow-up where one thing about the setup changes. Try the follow-up.