Zurück zu den Modellen
Forecaster

DirectReductionForecaster

API-ReferenzQuellcode
Categorical in XInsamplePred int insampleExogenous

Direct reduction forecaster, incl single-output, multi-output, exogenous Dir.

Implements direct reduction, of forecasting to tabular regression.

For no X, defaults to DirMO (direct multioutput) for X_treatment = "concurrent", and simple direct (direct single-output) for X_treatment = "shifted".

Direct single-output with concurrent X behaviour can be configured by passing a single-output scikit-learn compatible transformer.

Algorithm details:

In fit, given endogeneous time series y and possibly exogenous X: fits estimator to feature-label pairs as defined as follows.

  • if X_treatment = "concurrent": features = y(t), y(t-1), …, y(t-window_size), if provided X(t+h); labels = y(t+h) for h in the forecasting horizon, ranging over all t where the above have been observed (are in the index) for each h in the forecasting horizon (separate estimator fitted per h)

  • if X_treatment = "shifted": features = y(t), y(t-1), …, y(t-window_size), if provided: X(t); labels = y(t+h_1), …, y(t+h_k) for h_j in the forecasting horizon ranging over all t where the above have been observed (are in the index) estimator is fitted as a multi-output estimator (for all h_j simultaneously)

In predict, given possibly exogenous X, at cutoff time c,

  • if X_treatment = "concurrent": applies fitted estimators’ predict to feature = y(c), y(c-1), …, y(c-window_size), if provided X(c+h); to obtain a prediction for y(c+h), for each h in the forecasting horizon

  • if X_treatment = "shifted": applies fitted estimator’s predict to features = y(c), y(c-1), …, y(c-window_size), if provided: X(c); to obtain prediction for y(c+h_1), …, y(c+h_k) for h_j in forecasting horizon

Schnellstart

python
from sktime.forecasting.compose import DirectReductionForecaster

estimator = DirectReductionForecaster(estimator, window_length=10, transformers=None, X_treatment='concurrent', impute_method='bfill', pooling='local', windows_identical=False)

Parameter(7)

estimatorsklearn regressor, must be compatible with sklearn interface
tabular regression algorithm used in reduction algorithm
window_lengthint, optional, default=10
window length used in the reduction algorithm
transformerscurrently not used
X_treatmentstr, optional, one of “concurrent” (default) or “shifted”

determines the timestamps of X from which y(t+h) is predicted, for horizon h

  • “concurrent”: y(t+h) is predicted from lagged y, and X(t+h), for all h in fh in particular, if no y-lags are specified, y(t+h) is predicted from X(t)

  • “shifted”: y(t+h) is predicted from lagged y, and X(t), for all h in fh in particular, if no y-lags are specified, y(t+h) is predicted from X(t+h)

impute_methodstr, None, or sktime transformation, optional

Imputation method to use for missing values in the lagged data

  • default=”bfill”

  • if str, admissible strings are of Imputer.method parameter, see there. To pass further parameters, pass the Imputer transformer directly, as described below.

  • if sktime transformer, this transformer is applied to the lagged data. This needs to be a transformer that removes missing data, and can be an Imputer.

  • if None, no imputation is done when applying Lag transformer

poolingstr, one of [“local”, “global”, “panel”], optional, default=”local”

level on which data are pooled to fit the supervised regression model

  • “local” = unit/instance level, one reduced model per lowest hierarchy level

  • “global” = top level, one reduced model overall, on pooled data ignoring levels

  • “panel” = second lowest level, one reduced model per panel level (-2)

if there are 2 or less levels, “global” and “panel” result in the same if there is only 1 level (single time series), all three settings agree

windows_identicalbool, optional, default=False

Specifies whether all direct models use the same number of observations or a different number of observations.

  • True: Uniform window of length (total observations - maximum forecasting horizon). Note: Currently, there are no missing arising from window length due to backwards imputation in ReductionTransformer. Without imputation, the window size corresponds to (total observations + 1 - window_length + maximum forecasting horizon).

  • False: Window size differs for each forecasting horizon. Window length corresponds to (total observations + 1 - window_length + forecasting horizon).