Add A Custom Preprocessor

Custom preprocessors run inside the same split discipline as built-in preprocessing.

import macroforecast as mf

@mf.custom_preprocessor("center_x")
def center_x(X_train, y_train, X_test, context):
    location = X_train.mean(axis=0)
    return X_train - location, X_test - location

Use the registered name in an experiment:

result = (
    mf.Experiment(
        dataset="fred_md",
        target="INDPRO",
        start="1980-01",
        end="2019-12",
        horizons=[1, 3, 6],
        model_family="ridge",
    )
    .use_preprocessor("center_x")
    .run()
)

MVP contract:

fn(X_train, y_train, X_test, context) -> (X_train_new, X_test_new)

Rules:

• Fit preprocessing decisions on X_train only.

• Return transformed X_train and X_test.

• Do not transform y_train.

• Use y_train only as read-only context for target-aware feature preprocessing.

• Keep the number of training rows aligned with y_train.

Use a fixed custom preprocessor while comparing models:

result = (
    mf.Experiment(
        dataset="fred_md",
        target="INDPRO",
        start="1980-01",
        end="2019-12",
        model_family="ridge",
    )
    .use_preprocessor("center_x")
    .compare_models(["ridge", "lasso"])
    .run()
)

Preprocessor sweeps are blocked in the simple runtime until the preprocessing layer audit is complete. Target transformations are a separate contract. The MVP preprocessor API is predictor-only by design.