Temporal Cross-Validation

A random fold can leak information when financial labels use future prices or overlap in time. The notebooks compare random, ordered, purged, embargoed, and walk-forward designs on the same prediction task.

The central relation

Overlap between two information intervals is recorded by

I_i=[s_i,e_i],\qquad O_{ij}=\mathbf 1\!\left\{I_i\cap I_j\neq\varnothing\right\}.

$I_i$ is the information interval for observation $i$ , with start $s_i$ and end $e_i$ . The indicator $O_{ij}$ equals one when observations $i$ and $j$ share any part of their information intervals.

Keeping future information out of validation

The anatomy figure shows which observations are removed around a test fold. The public comparison reports held-out accuracy across several temporal designs.

Training, test, purged, and embargoed observations arranged in time order. — Purging removes training labels whose information overlaps the test period. The embargo removes a short interval immediately after the test fold.

Public-data accuracy across random, ordered, purged, and embargoed validation folds. — Fold-level results reveal how the validation estimate changes when time order and information overlap are handled explicitly.

Temporal protection reduces a known source of leakage. It does not make later market conditions identical to the evaluation sample.

Notebooks


1. Why Random Folds Fail for Overlapping Financial Labels
2. Preventing leakage between training and validation
3. Respecting time order in walk-forward validation