Member Junction
    Preparing search index...

    Validation discipline for a pipeline (plan §4.2 / §8.2). Defaults to a single train/test split with overfitting detection; k-fold and holdout are opt-in. A locked final holdout the search never sees is scored exactly once on the promoted model to produce the honest number.

    interface ValidationStrategy {
        K?: number;
        LockedHoldoutFraction?: number;
        MajorityBaselineRule?: MajorityBaselineRule;
        Strategy: "holdout" | "train_test_split" | "kfold";
        TestSize?: number;
        TimeIndexColumn?: string;
    }
    Index

    Properties

    K?: number

    Number of folds for kfold.

    LockedHoldoutFraction?: number

    The holdout share of a pipeline saved before Evaluation Protocols. Only the legacy-pipeline migration reads it (LegacyPipelines.LegacyProtocolInputs), to give such a pipeline its protocol; a pipeline's bound Evaluation Protocol decides its holdout, so a pipeline built with one carries no fraction.

    MajorityBaselineRule?: MajorityBaselineRule

    How the trust verdict treats a classifier no better than always guessing the most common class (or one that predicts a single class). Absent = 'enforce'. Recorded in the model's Lineage.evaluation at training time. To be replaced by the Evaluation Protocol.

    Strategy: "holdout" | "train_test_split" | "kfold"

    The validation approach.

    TestSize?: number

    Test fraction for train_test_split.

    TimeIndexColumn?: string

    Column that orders the rows in time (e.g. MonthIndex) when the target is a forecast. Declaring it tells Predictive Studio the model must be tested on later periods than it trained on; until that evaluation exists (a rolling-origin backtest) the trust verdict is withheld with the stored reason "not evaluated as a forecast". To be replaced by the Prediction Unit's Temporal Spec.