Sweep Only What You Care About

Use .sweep() when one part of the design should vary and everything else should stay on the default profile.

The executable MVP sweep is model comparison:

import macroforecast as mf

result = (
    mf.Experiment(
        dataset="fred_md",
        target="INDPRO",
        start="1980-01",
        end="2019-12",
        horizons=[1, 3, 6],
    )
    .sweep({"models": ["ar", "ridge", "lasso"]})
    .run()
)

Then rank the variants:

ranking = result.ranking  # or result.mean(metric="mse")

MVP aliases:

Alias	Internal axis	Runtime status
model, models, model_family	4_forecasting_model.nodes.fit_main.params.family	executable
scaling, scaling_policy	2_preprocessing.scaling_policy	full-route executable; simple exposure pending result-contract docs
missing, x_missing, x_missing_policy	2_preprocessing.x_missing_policy	full-route executable; simple exposure pending result-contract docs
preprocessor, custom_preprocessor	2_preprocessing.custom_preprocessor	fixed custom preprocessor executable; sweep blocked
target_transformer	3_feature_engineering.target_construction	fixed transformer executable for autoreg path; sweep blocked

A single-value choice becomes a fixed axis. A multi-value choice becomes a sweep axis.

exp = mf.Experiment(
    dataset="fred_md",
    target="INDPRO",
    start="1980-01",
    end="2019-12",
).sweep({"models": ["ridge"]})

recipe = exp.to_recipe_dict()

Representation and preprocessing sweeps are a full-route feature first. The full layer contract can route Layer 2 x Layer 3 x Layer 4 grids to the sweep runner. The simple API will expose this only after the public naming and result-summary contract is stable.

Full grids, ablations, replications, benchmark suites, and multi-target wrappers follow the same rule: they are internal/full routes first, simple routes later.