Custom Window, Selection, and Forecasting

Back to custom extensions

Use these hooks when the timing, parameter search, or forecast-combination logic is project-specific.

custom_stage_policy

mf.window.custom_stage_policy(selector, *, name=None, metadata=None) -> mf.window.StagePolicy

Selector Contract

selector(index: pandas.Index, *, item: dict, policy: StagePolicy)

The selector may return a boolean mask, a slice, integer positions, or index labels. The result must select at least one label and must not select dates outside the supplied index.

Example

def last_fit_half(index, *, item, policy):
    fit_idx = item["fit_idx"]
    return index[fit_idx[len(fit_idx) // 2 :]]

policy = mf.window.custom_stage_policy(last_fit_half)

Use this for non-standard model-selection or preprocessing samples. Standard expanding, rolling, fixed-reference, and origin-available behavior should use mf.window.stage_policy(...).

custom_search

mf.model_selection.custom_search(
    name,
    custom_func,
    **params,
) -> mf.model_selection.SearchSpec

Search Callable Contract

func(
    *,
    model,
    X,
    y,
    splits,
    metric,
    fixed_params,
    search,
    rng,
    maximize,
    evaluate_candidate,
    **params,
)

The callable may return trial records, a trial DataFrame, a SearchResult, or (records, metadata). Prefer evaluate_candidate(...) so the package owns fit/predict/scoring consistency.

custom_combination

mf.forecasting.custom_combination(
    name,
    func,
    *,
    models=None,
    **params,
) -> mf.forecasting.CombinationSpec

Combination Callable Contract

func(forecasts: pandas.DataFrame, *, actual: pandas.Series, **params)

forecasts is a wide matrix of base-model predictions indexed by (date, origin, origin_pos, horizon). The output must be a Series or one-dimensional array-like object aligned to those rows.

Example

def blend(forecasts, *, actual, weight=0.5):
    return weight * forecasts.iloc[:, 0] + (1.0 - weight) * forecasts.iloc[:, -1]

combination = mf.forecasting.custom_combination(
    "ridge_lasso_blend",
    blend,
    models=["ridge", "lasso"],
    weight=0.25,
)

Boundary

Hook	Owns	Does not own
custom_stage_policy	sample labels for one stage	fitting or scoring
custom_search	hyperparameter search algorithm	model default search spaces
custom_combination	combining already produced forecasts	fit-time model ensembles