How should time series cross-validation be done? Why can't we use normal K-Fold?
Quant ResearchLevel: Intermediate
How should time series cross-validation be done? Why can't we use normal K-Fold?
Hints
- Which observations must remain independent during evaluation?
- Compare random, grouped and temporal splits. Where are hyperparameters selected?
Answer
For a forecasting deployment, train on earlier observations and test on later ones using rolling or expanding windows. Ordinary K-fold can train on future observations even without shuffling. Purge training labels whose outcome intervals overlap test labels; use a justified gap or embargo where information dependencies require it. The required exclusion depends on labeling, timestamps and feature construction, not a universal five-day rule. Fit preprocessing and hyperparameters only inside each training/validation sequence.
Source and licence: Quant Research Interview Notes — SoYuCry; English translation by photon2563 · MIT · answer: Quantoros editorial