HKUDS/Vibe-Trading · error · ValueError

Purge and embargo removed all training samples for fold {fol

Error message

Purge and embargo removed all training samples for fold {fold}

What it means

After removing training observations whose labels overlap the test block (purge) and those within the embargo window after it, no training rows remain for the current fold. The splitter refuses to yield an empty train set because fitting a model on it would crash or silently produce garbage scores.

Source

Thrown at agent/src/quantlib/crossvalidation.py:270

        label_ends = np.arange(n_samples)
    else:
        label_ends = _as_label_spans(label_end_times, n_samples)

    embargo_size = int(round(n_samples * embargo_fraction))
    boundaries = np.linspace(0, n_samples, n_folds + 1).astype(int)

    for fold in range(n_folds):
        start, stop = int(boundaries[fold]), int(boundaries[fold + 1])
        if stop <= start:
            continue
        test_mask = np.zeros(n_samples, dtype=bool)
        test_mask[start:stop] = True

        train, purged, embargoed = _apply_purge_and_embargo(
            label_ends, test_mask, embargo_size
        )
        if train.size == 0:
            raise ValueError(f"Purge and embargo removed all training samples for fold {fold}")
        yield Split(
            train=train,
            test=np.arange(start, stop),
            purged=purged,
            embargoed=embargoed,
            test_bounds=(start, stop - 1),
        )


def group_purged_kfold_splits(
    groups: Sequence[object] | pd.Series | np.ndarray,
    n_folds: int = 5,
    embargo_fraction: float = DEFAULT_EMBARGO_FRACTION,
) -> Iterator[Split]:
    """Purged and embargoed K-fold cross-validation for panel and multi-asset datasets.

    Groups all observations sharing a time group identifier (e.g. date or bar timestamp)
    so that simultaneous observations across different assets are never split across

View on GitHub (pinned to 80ffdda44c)

Solutions

1. Reduce embargo_fraction
2. Reduce n_folds so test blocks (and their purge footprints) are larger and fewer
3. Use a shorter label horizon or a longer sample so enough non-overlapping train rows exist
4. Switch to purged_walk_forward_splits which naturally accounts for chronological label spans

Example fix

// before
splits = list(purged_kfold_splits(X, n_folds=10, embargo_fraction=0.3, label_end_times=le))
// after
splits = list(purged_kfold_splits(X, n_folds=5, embargo_fraction=0.05, label_end_times=le))

embargo_size = int(embargo_fraction * len(X))
# rough sanity: widest label span + embargo must leave room in the smallest train region
assert (len(X) // n_folds) > (label_ends.max() - np.arange(len(label_ends))).max() + embargo_size, 'risk of empty train fold'

try:
    splits = list(purged_kfold_splits(X, n_folds=n, embargo_fraction=e, label_end_times=le))
except ValueError as err:
    if 'removed all training samples' in str(err):
        n, e = max(2, n // 2), e / 4  # relax and retry once
        splits = list(purged_kfold_splits(X, n_folds=n, embargo_fraction=e, label_end_times=le))
    else:
        raise

When it happens

Trigger: Long label horizons relative to the sample plus a large embargo_fraction: for a fold near the start or end of the data, every candidate training row overlaps a test label or falls inside the embargo window (e.g. label_end_times spanning 50% of the sample with 10 folds).

Common situations: Overlapping multi-day labels on short price history, high fold counts with wide labels, or an aggressive embargo_fraction copied from a longer dataset.

Related errors

• embargo_fraction must be in [0, 1), got {embargo_fraction}
• a label cannot end before the observation it belongs to star
• label_end_times has {span_ends.size} entries but the sample
• n_folds must be at least {MIN_FOLDS}, got {n_folds}
• {n_samples} samples cannot make {n_folds} folds

AI-assisted analysis of HKUDS/Vibe-Trading@80ffdda44c (2026-08-28). Data as JSON: /api/errors/73e8eed34fe677d7. Report an issue: GitHub.