Time-series model selection

Walk-forward, purged, k-fold and temporal splits for leakage-free time-series model selection.

model_selection provides cross-validation splitters that respect time — the foundation for validating any model trained on sequential data (markets, sensors, demand). Random k-fold leaks the future into training; these splitters never do.

All functions run natively in the Mojo engine (zero Python imports, deterministic, seeded where applicable).

The modules

Module	What it does
model_selection	leakage-free cross-validation splitters for sequential data — walk-forward, purged, k-fold and temporal splits that respect time ordering instead of leaking the future into training the way random k-fold does

Functions

Function	What it does
walk_forward_splits(n, n_splits, min_train, expanding)	Ordered walk-forward folds: train on [0, k), test on the next contiguous window. expanding=True grows the training window each fold; False rolls it at min_train. Test windows are contiguous, non-overlapping, and cover the tail.
purged_walk_forward_splits(n, n_splits, min_train, purge_gap, embargo)	Finance-grade walk-forward: drops purge_gap samples between train end and test start (removes label leakage from overlapping forward-return windows) and embargoes embargo samples after each test window from later training folds.
kfold_splits(n, k, seed, shuffle)	Classic k-fold for non-temporal data.
stratified_kfold_splits(labels, k, seed)	K-fold preserving class proportions per fold.
train_test_split_temporal(n, test_fraction)	Single ordered split.

When to use purged CV

If your label at time t is computed from data in (t, t+h] (a forward return, a settlement outcome), a plain walk-forward split lets training samples within h of the test boundary see part of the test window's future. Set purge_gap >= h to remove them. This is the standard Lopez de Prado correction for overlapping-label leakage.

Verification

Every splitter is covered by known-value geometry tests plus a 50k-split invariant fuzzer.

Next: promote what you trained with statistical model comparison, or browse the engine catalog.