Forecaster

ForecastX

Categorical in XInsamplePred intPred int insampleMissing values Exogenous Multivariate

Forecaster that forecasts exogenous data for use in an endogeneous forecast.

In predict, this forecaster carries out a predict step on exogenous X. Then, a forecast is made for y, using exogenous data plus its forecasts as X. If columns argument is provided, will carry predict out only for the columns in columns, and will use other columns in X unchanged.

The two forecasters and forecasting horizons (for forecasting y resp X) can be selected independently, but default to the same.

The typical use case is extending exogenous data available only up until the cutoff into the future, for use by an exogenous forecaster that requires such future data.

If no X is passed in fit, behaves like forecaster_y. In such a case (no exogenous data), there is no benefit in using this compositor.

If variables in columns are present in the provided X during predict, by default these are still forecasted and the forecasts are used for prediction of y variables. This behaviour can be modified by passing predict_behaviour argument as "use_actuals" instead of the default value of "use_forecasts".

Quickstart

python

from sktime.forecasting.compose import ForecastX

estimator = ForecastX(forecaster_y, forecaster_X=None, fh_X=None, behaviour='update', columns=None, fit_behaviour='use_actual', forecaster_X_exogeneous='None', predict_behaviour='use_forecasts')

Parameters(8)

forecaster_yBaseForecaster

sktime forecaster to use for endogeneous data y

forecaster_XBaseForecaster, optional

sktime forecaster to use for exogenous data X, default = None = same as forecaster_y

fh_XNone, ForecastingHorizon, or valid input to construct ForecastingHorizon

optional, default = None = same as used for y in any instance. valid inputs to construct ForecastingHorizon are: int, list of int, 1D np.ndarray, pandas.Index (see ForecastingHorizon)

behaviourstr, one of “update” or “refit”, optional, default = “update”

if “update”, forecaster_X is fit to the data batch seen in fit,

and updated with any X seen in calls of update. Forecast added to X in predict is obtained from this state.

if “refit”, then forecaster_X is fit to X in predict only,

Forecast added to X in predict is obtained from this state.

columnsNone, or pandas compatible index iterator (e.g., list of str), optional

default = None = all columns in X are used for forecast columns to which forecaster_X is applied. If not None, must be a non-empty list of valid column names. Note that [] and None do not imply the same.

fit_behaviourstr, one of “use_actual” (default), “use_forecast”, optional,

if “use_actual”, then forecaster_y uses the actual X as

exogenous features in fit * if “use_forecast”, then forecaster_y uses the X predicted by forecaster_X as exogenous features in fit

forecaster_X_exogeneousoptional, str, one of “None” (default), or “complement”,

or pandas.Index coercible

if “None”, then forecaster_X uses no exogenous data
if “complement”, then forecaster_X uses the complement of the

columns as exogenous data to forecast. This is typically useful if the complement of columns is known to be available in the future. * if a pandas.Index coercible, then uses columns indexed by the index after coercion, in X passed (converted to pandas)

predict_behaviourstr, optional (default = “use_forecasts”)

if “use_forecasts”, then forecaster_X predictions are always used as
inputs in forecaster_y, even if passed X has future values

Examples

>>> from sktime.datasets import load_longley
>>> from sktime.forecasting.arima import ARIMA
>>> from sktime.forecasting.base import ForecastingHorizon
>>> from sktime.forecasting.compose import ForecastX
>>> from sktime.forecasting.var import VAR
>>> y, X = load_longley ()
>>> fh = ForecastingHorizon ([1, 2, 3 ])
>>> pipe = ForecastX (
... forecaster_X = VAR (),
... forecaster_y = ARIMA (),
... )
>>> pipe = pipe. fit (y, X = X, fh = fh)
>>> # this now works without X from the future of y!
>>> y_pred = pipe. predict (fh = fh) to forecast only some columns, use the columns arg, and pass known columns to predict:
>>> columns = ["ARMED", "POP" ]
>>> pipe = ForecastX (
... forecaster_X = VAR (),
... forecaster_y = SARIMAX (),
... columns = columns,
... )
>>> pipe = pipe. fit (y_train, X = X_train, fh = fh)
>>> # dropping ["ARMED", "POP"] = columns where we expect not to have future values
>>> y_pred = pipe. predict (fh = fh, X = X_test. drop (columns = columns))