check_class_proba_prediction_sums¶

def check_class_proba_prediction_sums(forecaster, y_test: pl.DataFrame) -> None:
    """Check probabilities sum to 1.0 per row per target (tolerance: 1e-6).

    Parameters
    ----------
    forecaster : BaseClassProbaForecaster
        Fitted class-probability forecaster instance.
    y_test : pl.DataFrame
        Test target data.

    Raises
    ------
    AssertionError
        If probabilities do not sum to approximately 1.0.

    """
    forecasting_horizon = min(3, len(y_test))
    y_pred = forecaster.predict_class_proba(forecasting_horizon=forecasting_horizon)

    _, y_panel_groups = inspect_panel(y_test)

    if len(y_panel_groups) > 0:
        for group_prefix in y_panel_groups:
            for target_col, class_labels in forecaster.classes_.items():
                proba_cols = [f"{group_prefix}__{target_col}_proba_{label}" for label in class_labels]
                row_sums = y_pred.select(proba_cols).sum_horizontal()
                max_err = (row_sums - 1.0).abs().max()
                assert max_err < 1e-6, (
                    f"Probabilities for {group_prefix}__{target_col} deviate from 1.0 by up to {max_err:.2e}, expected < 1e-6"
                )
    else:
        for target_col, class_labels in forecaster.classes_.items():
            proba_cols = [f"{target_col}_proba_{label}" for label in class_labels]
            row_sums = y_pred.select(proba_cols).sum_horizontal()
            max_err = (row_sums - 1.0).abs().max()
            assert max_err < 1e-6, (
                f"Probabilities for {target_col} deviate from 1.0 by up to {max_err:.2e}, expected < 1e-6"
            )