GridSearchCV details¶

The shared wrapper is run_grid_search() in utils/train_utils.py.

Each AutoML model provides:

The wrapper handles cross validation, scoring, refit, and result formatting.

Cross validation¶

The project uses:

StratifiedKFold(
    n_splits=config.CV_FOLDS,
    shuffle=True,
    random_state=config.RANDOM_STATE,
)

config.CV_FOLDS is currently 3.

StratifiedKFold keeps the class ratio similar across folds, which is useful for binary risk labels.

SCORING = {
    "f1": "f1",
    "roc_auc": "roc_auc",
    "accuracy": "accuracy",
}

The best parameter set is selected with:

REFIT_METRIC = "roc_auc"

So best_params_ means “best by mean validation ROC AUC”, not best by F1 or Accuracy.

Grid search tries every parameter combination. A Random Forest grid with 24 combinations and 3 folds runs 72 model fits.

Several estimators are configured with n_jobs=1 internally because the outer GridSearchCV already parallelizes with n_jobs=-1.