ForecasterRecursiveMultiSeries

skforecast.recursive._forecaster_recursive_multiseries.ForecasterRecursiveMultiSeries

ForecasterRecursiveMultiSeries(
    estimator,
    lags=None,
    window_features=None,
    calendar_features=None,
    encoding="ordinal",
    transformer_series=None,
    transformer_exog=None,
    categorical_features="auto",
    weight_func=None,
    series_weights=None,
    differentiation=None,
    dropna_from_series=False,
    fit_kwargs=None,
    binner_kwargs=None,
    forecaster_id=None,
)

Bases: ForecasterBase

This class turns any estimator compatible with the scikit-learn API into a recursive autoregressive (multi-step) forecaster for multiple series.

Parameters:

Name	Type	Description	Default
estimator	estimator or pipeline compatible with the scikit-learn API	An instance of an estimator or pipeline compatible with the scikit-learn API.	required
lags	int, list, numpy ndarray, range	Lags used as predictors. Index starts at 1, so lag 1 is equal to t-1. int: include lags from 1 to lags (included). list, 1d numpy ndarray or range: include only lags present in lags, all elements must be int. None: no lags are included as predictors.	None
window_features	(object, list)	Instance or list of instances used to create window features. Window features are created from the original time series and are included as predictors. Skforecast provides the RollingFeatures class, but a custom object can also be passed as long as it implements the required interface.	None
calendar_features	object	Instance of CalendarFeatures used to create calendar features from the datetime index. Calendar features are included as predictors and are generated automatically during both training and prediction. Only supported when the index of the input data is a pandas.DatetimeIndex. New in version 0.23.0	None
encoding	(str, None)	Encoding used to identify the different series. If 'ordinal', a single column is created with integer values from 0 to n_series - 1. If 'ordinal_category', a single column is created with integer values from 0 to n_series - 1 and the column is transformed into pandas.category dtype so that it can be used as a categorical variable. If 'onehot', a binary column is created for each series. If None, no column is created to identify the series. Internally, the series are identified as an integer from 0 to n_series - 1, but no column is created in the training matrices.	'ordinal'
transformer_series	(transformer(preprocessor), dict)	An instance of a transformer (preprocessor) compatible with the scikit-learn preprocessing API with methods: fit, transform, fit_transform and inverse_transform. Transformation is applied to each series before training the forecaster. ColumnTransformers are not allowed since they do not have inverse_transform method. If single transformer: it is cloned and applied to all series. If dict of transformers: a different transformer can be used for each series.	None
transformer_exog	transformer	An instance of a transformer (preprocessor) compatible with the scikit-learn preprocessing API. The transformation is applied to exog before training the forecaster. inverse_transform is not available when using ColumnTransformers.	None
categorical_features	(str, list)	Specifies which exogenous variables should be treated as categorical features. Categorical features are encoded using an OrdinalEncoder internally managed by the forecaster. If 'auto': after applying transformer_exog, any column with a non-numeric dtype is treated as categorical. If list: a list of column names to be treated as categorical. If None: no categorical encoding is applied internally. New in version 0.22.0	'auto'
weight_func	(Callable, dict)	Function that defines the individual weights for each sample based on the index. For example, a function that assigns a lower weight to certain dates. Ignored if estimator does not have the argument sample_weight in its fit method. See Notes section for more details on the use of the weights. If single function: it is applied to all series. If dict {'series_column_name' : Callable}: a different function can be used for each series, a weight of 1 is given to all series not present in weight_func.	None
series_weights	dict	Weights associated with each series {'series_column_name' : float}. It is only applied if the estimator used accepts sample_weight in its fit method. See Notes section for more details on the use of the weights. If a dict is provided, a weight of 1 is given to all series not present in series_weights. If None, all levels have the same weight.	None
differentiation	(int, dict)	Order of differencing applied to the time series before training the forecaster. The order of differentiation is the number of times the differencing operation is applied to a time series. Differencing involves computing the differences between consecutive data points in the series. Before returning a prediction, the differencing operation is reversed. If int, the same order of differentiation is applied to all series. If dict, a different order of differentiation (including None) can be used for each series. The keys must be the names of the series used to fit the forecaster. If a series is not present in the dictionary, no differencing is applied. If None, no differencing is applied.	None
dropna_from_series	bool	Determine whether NaN detected in the training matrices will be dropped. Relevant when series or exog contain interspersed NaN values. If True, drop NaNs in X_train and same rows in y_train. If False, leave NaNs in X_train and warn the user.	False
fit_kwargs	dict	Additional arguments to be passed to the fit method of the estimator.	None
binner_kwargs	dict	Additional arguments to pass to the QuantileBinner used to discretize the residuals into k bins according to the predicted values associated with each residual. Available arguments are: n_bins, method, subsample, random_state and dtype. Argument method is passed internally to the function numpy.percentile. New in version 0.14.0	None
forecaster_id	(str, int)	Name used as an identifier of the forecaster.	None

Attributes:

Name	Type	Description
estimator	estimator or pipeline compatible with the scikit-learn API	An instance of an estimator or pipeline compatible with the scikit-learn API.
lags	numpy ndarray	Lags used as predictors.
lags_names	list	Names of the lags used as predictors.
max_lag	int	Maximum lag included in lags.
window_features	list	Class or list of classes used to create window features.
window_features_names	list	Names of the window features to be included in the X_train matrix.
window_features_class_names	list	Names of the classes used to create the window features.
max_size_window_features	int	Maximum window size required by the window features.
calendar_features	object	Instance of CalendarFeatures used to create calendar features from the datetime index.
calendar_features_names	list	Names of the calendar features to extract, taken from the features attribute of the calendar_features object.
window_size	int	The window size needed to create the predictors. It is calculated as the maximum value between max_lag and max_size_window_features. If differentiation is used, window_size is increased by n units equal to the order of differentiation so that predictors can be generated correctly.
encoding	str	Encoding used to identify the different series. If 'ordinal', a single column is created with integer values from 0 to n_series - 1. If 'ordinal_category', a single column is created with integer values from 0 to n_series - 1 and the column is transformed into pandas.category dtype so that it can be used as a categorical variable. If 'onehot', a binary column is created for each series. If None, no column is created to identify the series. Internally, the series are identified as an integer from 0 to n_series - 1, but no column is created in the training matrices.
encoding_mapping_	dict	Mapping of the encoding used to identify the different series.
transformer_series	(transformer(preprocessor), dict)	An instance of a transformer (preprocessor) compatible with the scikit-learn preprocessing API with methods: fit, transform, fit_transform and inverse_transform. Transformation is applied to each series before training the forecaster. ColumnTransformers are not allowed since they do not have inverse_transform method. If single transformer: it is cloned and applied to all series. If dict of transformers: a different transformer can be used for each series.
transformer_series_	dict	Dictionary with the transformer for each series. It is created cloning the objects in transformer_series and is used internally to avoid overwriting.
transformer_exog	transformer(preprocessor)	An instance of a transformer (preprocessor) compatible with the scikit-learn preprocessing API. The transformation is applied to exog before training the forecaster. inverse_transform is not available when using ColumnTransformers.
weight_func	(Callable, dict)	Function that defines the individual weights for each sample based on the index. For example, a function that assigns a lower weight to certain dates. Ignored if estimator does not have the argument sample_weight in its fit method. See Notes section for more details on the use of the weights. If single function: it is applied to all series. If dict {'series_column_name' : Callable}: a different function can be used for each series, a weight of 1 is given to all series not present in weight_func.
weight_func_	dict	Dictionary with the weight_func for each series. It is created cloning the objects in weight_func and is used internally to avoid overwriting.
source_code_weight_func	(str, dict)	Source code of the custom function(s) used to create weights.
series_weights	dict	Weights associated with each series {'series_column_name' : float}. It is only applied if the estimator used accepts sample_weight in its fit method. See Notes section for more details on the use of the weights. If a dict is provided, a weight of 1 is given to all series not present in series_weights. If None, all levels have the same weight.
series_weights_	dict	Weights associated with each series. It is created as a clone of series_weights and is used internally to avoid overwriting.
differentiation	(int, dict)	The order of differencing applied to the time series prior to training the forecaster.
differentiation_max	int	Maximum order of differentiation.
differentiator	(TimeSeriesDifferentiator, dict)	Skforecast object (or dict of objects) used to differentiate the time series.
differentiator_	dict	Dictionary with the differentiator for each series. It is created cloning the objects in differentiator and is used internally to avoid overwriting.
dropna_from_series	bool	Determine whether NaN detected in the training matrices will be dropped.
last_window_	dict	Last window of training data for each series. It stores the values needed to predict the next step immediately after the training data. These values are stored in the original scale of the time series before undergoing any transformations or differentiation. When differentiation parameter is specified, the dimensions of the last_window_ are expanded as many values as the order of differentiation. For example, if lags = 7 and differentiation = 1, last_window_ will have 8 values.
index_type_	type	Type of index of the input used in training.
index_freq_	str	Frequency of Index of the input used in training.
training_range_	dict	First and last values of index of the data used during training for each series.
series_names_in_	list	Names of the series (levels) provided by the user during training.
exog_in_	bool	If the forecaster has been trained using exogenous variable/s.
exog_names_in_	list	Names of the exogenous variables used during training.
exog_type_in_	type	Type of exogenous data (pandas Series, DataFrame or dict) used in training.
exog_dtypes_in_	dict	Type of each exogenous variable/s used in training before the transformation applied by transformer_exog. If transformer_exog is not used, it is equal to exog_dtypes_out_.
exog_dtypes_out_	dict	Type of each exogenous variable/s used in training after the transformation applied by transformer_exog. If transformer_exog is not used, it is equal to exog_dtypes_in_.
X_train_series_names_in_	list	Names of the series (levels) included in the matrix X_train created internally for training. It can be different from series_names_in_ if some series are dropped during the training process because of NaNs or because they are not present in the training period.
X_train_window_features_names_out_	list	Names of the window features included in the matrix X_train created internally for training.
X_train_calendar_features_names_out_	list	Names of the calendar features included in the matrix X_train created internally for training.
X_train_exog_names_out_	list	Names of the exogenous variables included in the matrix X_train created internally for training. It can be different from exog_names_in_ if some exogenous variables are transformed during the training process.
X_train_features_names_out_	list	Names of columns of the matrix created internally for training.
categorical_features	(str, list)	How categorical features are identified among the exogenous variables. It can be 'auto', a list of column names or None.
categorical_features_names_in_	list	Names of the exogenous variables considered as categorical.
categorical_encoder	sklearn OrdinalEncoder	OrdinalEncoder used internally to encode categorical features.
fit_kwargs	dict	Additional arguments to be passed to the fit method of the estimator.
in_sample_residuals_	dict	Residuals of the model when predicting training data. Only stored up to 10_000 values per series in the form {series: residuals}. If transformer_series is not None, residuals are stored in the transformed scale. If differentiation is not None, residuals are stored after differentiation.
in_sample_residuals_by_bin_	dict	In-sample residuals binned according to the predicted value each residual is associated with. The number of residuals stored per bin is limited to 10_000 // self.binner.n_bins_ per series in the form {series: residuals}. If transformer_series is not None, residuals are stored in the transformed scale. If differentiation is not None, residuals are stored after differentiation.
out_sample_residuals_	dict	Residuals of the model when predicting non-training data. Only stored up to 10_000 values per series in the form {series: residuals}. Use set_out_sample_residuals() method to set values. If transformer_series is not None, residuals are stored in the transformed scale. If differentiation is not None, residuals are stored after differentiation.
out_sample_residuals_by_bin_	dict	Out of sample residuals binned according to the predicted value each residual is associated with. The number of residuals stored per bin is limited to 10_000 // self.binner.n_bins_ per series in the form {series: residuals}. If transformer_series is not None, residuals are stored in the transformed scale. If differentiation is not None, residuals are stored after differentiation.
binner	dict	Dictionary of skforecast.preprocessing.QuantileBinner used to discretize residuals of each series into k bins according to the predicted values associated with each residual. In the form {series: binner}.
binner_intervals_	dict	Intervals used to discretize residuals into k bins according to the predicted values associated with each residual. In the form {series: binner_intervals_}.
binner_kwargs	dict	Additional arguments to pass to the QuantileBinner.
creation_date	str	Date of creation.
is_fitted	bool	Tag to identify if the estimator has been fitted (trained).
fit_date	str	Date of last fit.
skforecast_version	str	Version of skforecast library used to create the forecaster.
python_version	str	Version of python used to create the forecaster.
forecaster_id	(str, int)	Name used as an identifier of the forecaster.
__skforecast_tags__	dict	Tags associated with the forecaster.
_probabilistic_mode	(str, bool)	Private attribute used to indicate whether the forecaster should perform some calculations during backtesting.

Notes

The weights are used to control the influence that each observation has on the training of the model. ForecasterRecursiveMultiSeries accepts two types of weights. If the two types of weights are indicated, they are multiplied to create the final weights. The resulting sample_weight cannot have negative values.

• series_weights : controls the relative importance of each series. If a series has twice as much weight as the others, the observations of that series influence the training twice as much. The higher the weight of a series relative to the others, the more the model will focus on trying to learn that series.
• weight_func : controls the relative importance of each observation according to its index value. For example, a function that assigns a lower weight to certain dates.

Methods:

Name	Description
create_train_X_y	Create training matrices from multiple time series and exogenous
create_sample_weights	Create weights for each observation according to the forecaster's attributes
fit	Training Forecaster. See Notes section for more details depending on
create_predict_X	Create the predictors needed to predict steps ahead. As it is a recursive
predict	Predict n steps ahead. It is a recursive process in which, each prediction,
predict_bootstrapping	Generate multiple forecasting predictions using a bootstrapping process.
predict_interval	Predict n steps ahead and estimate prediction intervals using either
predict_quantiles	Calculate the specified quantiles for each step. After generating
predict_dist	Fit a given probability distribution for each step. After generating
set_params	Set new values to the parameters of the scikit-learn model stored in the
set_lags	Set new value to the attribute lags. Attributes lags_names,
set_window_features	Set new value to the attribute window_features. Attributes
set_fit_kwargs	Set new values for the additional keyword arguments passed to the fit
set_in_sample_residuals	Set in-sample residuals in case they were not calculated during the
set_out_sample_residuals	Set new values to the attribute out_sample_residuals_. Out of sample
get_feature_importances	Return feature importances of the estimator stored in the

Source code in skforecast/recursive/_forecaster_recursive_multiseries.py

def __init__(
    self,
    estimator: object,
    lags: int | list[int] | np.ndarray[int] | range[int] | None = None,
    window_features: object | list[object] | None = None,
    calendar_features: object | None = None,
    encoding: str | None = 'ordinal',
    transformer_series: object | dict[str, object] | None = None,
    transformer_exog: object | None = None,
    categorical_features: str | list[str] | None = 'auto',
    weight_func: Callable | dict[str, Callable] | None = None,
    series_weights: dict[str, float] | None = None,
    differentiation: int | dict[str, int | None] | None = None,
    dropna_from_series: bool = False,
    fit_kwargs: dict[str, object] | None = None,
    binner_kwargs: dict[str, object] | None = None,
    forecaster_id: str | int | None = None
) -> None:

    self.estimator                            = clone(estimator)
    self.calendar_features                    = (
        clone(calendar_features) if calendar_features is not None else None
    )
    self.calendar_features_names              = getattr(calendar_features, 'features', None)
    self.encoding                             = encoding
    self.encoding_mapping_                    = {}
    self.transformer_series                   = transformer_series
    self.transformer_series_                  = None
    self.transformer_exog                     = transformer_exog
    self.categorical_features                 = categorical_features
    self.weight_func                          = weight_func
    self.weight_func_                         = None
    self.source_code_weight_func              = None
    self.series_weights                       = series_weights
    self.series_weights_                      = None
    self.differentiation                      = differentiation
    self.differentiation_max                  = None
    self.differentiator                       = None
    self.differentiator_                      = None
    self.dropna_from_series                   = dropna_from_series
    self.last_window_                         = None
    self.index_type_                          = None
    self.index_freq_                          = None
    self.training_range_                      = None
    self.series_names_in_                     = None
    self.exog_in_                             = False
    self.exog_names_in_                       = None
    self.exog_type_in_                        = None
    self.exog_dtypes_in_                      = None
    self.exog_dtypes_out_                     = None
    self.categorical_features_names_in_       = None
    self.X_train_series_names_in_             = None
    self.X_train_window_features_names_out_   = None
    self.X_train_calendar_features_names_out_ = None
    self.X_train_exog_names_out_              = None
    self.X_train_features_names_out_          = None
    self.in_sample_residuals_                 = None
    self.in_sample_residuals_by_bin_          = None
    self.out_sample_residuals_                = None
    self.out_sample_residuals_by_bin_         = None
    self.creation_date                        = pd.Timestamp.today().strftime('%Y-%m-%d %H:%M:%S')
    self.is_fitted                            = False
    self.fit_date                             = None
    self.skforecast_version                   = __version__
    self.python_version                       = sys.version.split(" ")[0]
    self.forecaster_id                        = forecaster_id
    self._probabilistic_mode                  = "binned"

    self.lags, self.lags_names, self.max_lag = initialize_lags(type(self).__name__, lags)
    self.lags_are_contiguous = (
        self.lags is not None
        and np.array_equal(self.lags, np.arange(1, self.max_lag + 1))
    )
    self.window_features, self.window_features_names, self.max_size_window_features = (
        initialize_window_features(window_features)
    )
    if self.window_features is None and self.lags is None:
        raise ValueError(
            "At least one of the arguments `lags` or `window_features` "
            "must be different from None. This is required to create the "
            "predictors used in training the forecaster."
        )

    self.window_size = max(
        [ws for ws in [self.max_lag, self.max_size_window_features] 
         if ws is not None]
    )
    self.window_features_class_names = None
    if window_features is not None:
        self.window_features_class_names = [
            type(wf).__name__ for wf in self.window_features
        ]

    if categorical_features is not None:
        if not (
            (isinstance(categorical_features, str) and categorical_features == 'auto')
            or isinstance(categorical_features, list)
        ):
            raise ValueError(
                f"Argument `categorical_features` must be `'auto'`, a list of "
                f"column names, or `None`. Got {categorical_features}."
            )
        if isinstance(categorical_features, list):
            if len(categorical_features) == 0:
                raise ValueError(
                    "Argument `categorical_features` must not be an empty list. "
                    "Use `None` to disable categorical encoding."
                )

    self.categorical_encoder = OrdinalEncoder(
                                   dtype                 = float,
                                   handle_unknown        = 'use_encoded_value',
                                   unknown_value         = np.nan,
                                   encoded_missing_value = np.nan
                               ).set_output(transform="pandas")

    if self.encoding not in ['ordinal', 'ordinal_category', 'onehot', None]:
        raise ValueError(
            f"Argument `encoding` must be one of the following values: 'ordinal', "
            f"'ordinal_category', 'onehot' or None. Got '{self.encoding}'."
        )

    if self.transformer_series is None and isinstance(estimator, (LinearModel, BaseLibSVM)):
        warnings.warn(
            "When using a linear model, it is recommended to use a transformer_series "
            "to ensure all series are in the same scale. You can use, for example, a "
            "`StandardScaler` from sklearn.preprocessing.",
            DataTransformationWarning
        )

    if isinstance(self.transformer_series, dict):
        if self.encoding is None:
            raise TypeError(
                "When `encoding` is None, `transformer_series` must be a single "
                "transformer (not `dict`) as it is applied to all series."
            )
        if '_unknown_level' not in self.transformer_series.keys():
            raise ValueError(
                "If `transformer_series` is a `dict`, a transformer must be "
                "provided to transform series that do not exist during training. "
                "Add the key '_unknown_level' to `transformer_series`. "
                "For example: {'_unknown_level': your_transformer}."
            )

    self.weight_func, self.source_code_weight_func, self.series_weights = (
        initialize_weights(
            forecaster_name = type(self).__name__,
            estimator       = estimator,
            weight_func     = weight_func,
            series_weights  = series_weights,
        )
    )

    if differentiation is not None:
        if isinstance(differentiation, int):
            if differentiation < 1:
                raise ValueError(
                    f"If `differentiation` is an integer, it must be equal "
                    f"to or greater than 1. Got {differentiation}."
                )
            self.differentiation = differentiation
            self.differentiation_max = differentiation
            self.window_size += self.differentiation_max
            self.differentiator = TimeSeriesDifferentiator(
                order=differentiation, window_size=self.window_size
            )
        elif isinstance(differentiation, dict):

            if self.encoding is None:
                raise TypeError(
                    "When `encoding` is None, `differentiation` must be an "
                    "integer equal to or greater than 1. Same differentiation "
                    "must be applied to all series."
                )
            if '_unknown_level' not in differentiation.keys():
                raise ValueError(
                    "If `differentiation` is a `dict`, an order must be provided "
                    "to differentiate series that do not exist during training. "
                    "Add the key '_unknown_level' to `differentiation`. "
                    "For example: {'_unknown_level': 1}."
                )

            differentiation_max = []
            for level, diff in differentiation.items():
                if diff is not None:
                    if not isinstance(diff, int) or diff < 1:
                        raise ValueError(
                            f"If `differentiation` is a dict, the values must be "
                            f"None or integers equal to or greater than 1. "
                            f"Got {diff} for series '{level}'."
                        )
                    differentiation_max.append(diff)

            if len(differentiation_max) == 0:
                raise ValueError(
                    "If `differentiation` is a dict, at least one value must be "
                    "different from None. Got all values equal to None. If you "
                    "do not want to differentiate any series, set `differentiation` "
                    "to None."
                )

            self.differentiation = differentiation
            self.differentiation_max = max(differentiation_max)
            self.window_size += self.differentiation_max
            self.differentiator = {
                level: (
                    TimeSeriesDifferentiator(order=diff, window_size=self.window_size)
                    if diff is not None else None
                )
                for level, diff in differentiation.items()
            }
        else:
            raise TypeError(
                f"When including `differentiation`, this argument must be "
                f"an integer (equal to or greater than 1) or a dict of "
                f"integers. Got {type(differentiation)}."
            )

    self.fit_kwargs = check_select_fit_kwargs(
                          estimator  = estimator,
                          fit_kwargs = fit_kwargs
                      )

    self.binner = {}
    self.binner_intervals_ = {}
    self.binner_kwargs = binner_kwargs
    if binner_kwargs is None:
        self.binner_kwargs = {
            'n_bins': 10, 'method': 'linear', 'subsample': 200000,
            'random_state': 789654, 'dtype': np.float64
        }

    self.__skforecast_tags__ = {
        "library": "skforecast",
        "forecaster_name": "ForecasterRecursiveMultiSeries",
        "forecaster_task": "regression",
        "forecasting_scope": "global",  # single-series | global
        "forecasting_strategy": "recursive",  # recursive | direct | deep_learning | foundation
        "multiple_estimators": False, 
        "index_types_supported": ["pandas.RangeIndex", "pandas.DatetimeIndex"],
        "requires_index_frequency": True,

        "allowed_input_types_series": ["pandas.DataFrame", "dict"],
        "supports_exog": True,
        "allowed_input_types_exog": ["pandas.Series", "pandas.DataFrame", "dict"],
        "handles_missing_values_series": True, 
        "handles_missing_values_exog": True, 

        "supports_lags": True,
        "supports_window_features": True,
        "supports_calendar_features": True,
        "allowed_encoding": ["ordinal", "ordinal_category", "onehot", None],
        "supports_transformer_series": True,
        "supports_transformer_exog": True,
        "supports_categorical_features": True,
        "supports_weight_func": True,
        "supports_series_weights": True,
        "supports_differentiation": True,

        "prediction_types": ["point", "interval", "bootstrapping", "quantiles", "distribution"],
        "supports_probabilistic": True,
        "probabilistic_methods": ["bootstrapping", "conformal"],
        "handles_binned_residuals": True
    }

Attributes

estimator instance-attribute

estimator = clone(estimator)

calendar_features instance-attribute

calendar_features = (
    clone(calendar_features)
    if calendar_features is not None
    else None
)

calendar_features_names instance-attribute

calendar_features_names = getattr(
    calendar_features, "features", None
)

encoding instance-attribute

encoding = encoding

encoding_mapping_ instance-attribute

encoding_mapping_ = {}

transformer_series instance-attribute

transformer_series = transformer_series

transformer_series_ instance-attribute

transformer_series_ = None

transformer_exog instance-attribute

transformer_exog = transformer_exog

categorical_features instance-attribute

categorical_features = categorical_features

weight_func instance-attribute

weight_func = weight_func

weight_func_ instance-attribute

weight_func_ = None

source_code_weight_func instance-attribute

source_code_weight_func = None

series_weights instance-attribute

series_weights = series_weights

series_weights_ instance-attribute

series_weights_ = None

differentiation instance-attribute

differentiation = differentiation

differentiation_max instance-attribute

differentiation_max = None

differentiator instance-attribute

differentiator = None

differentiator_ instance-attribute

differentiator_ = None

dropna_from_series instance-attribute

dropna_from_series = dropna_from_series

last_window_ instance-attribute

last_window_ = None

index_type_ instance-attribute

index_type_ = None

index_freq_ instance-attribute

index_freq_ = None

training_range_ instance-attribute

training_range_ = None

series_names_in_ instance-attribute

series_names_in_ = None

exog_in_ instance-attribute

exog_in_ = False

exog_names_in_ instance-attribute

exog_names_in_ = None

exog_type_in_ instance-attribute

exog_type_in_ = None

exog_dtypes_in_ instance-attribute

exog_dtypes_in_ = None

exog_dtypes_out_ instance-attribute

exog_dtypes_out_ = None

categorical_features_names_in_ instance-attribute

categorical_features_names_in_ = None

X_train_series_names_in_ instance-attribute

X_train_series_names_in_ = None

X_train_window_features_names_out_ instance-attribute

X_train_window_features_names_out_ = None

X_train_calendar_features_names_out_ instance-attribute

X_train_calendar_features_names_out_ = None

X_train_exog_names_out_ instance-attribute

X_train_exog_names_out_ = None

X_train_features_names_out_ instance-attribute

X_train_features_names_out_ = None

in_sample_residuals_ instance-attribute

in_sample_residuals_ = None

in_sample_residuals_by_bin_ instance-attribute

in_sample_residuals_by_bin_ = None

out_sample_residuals_ instance-attribute

out_sample_residuals_ = None

out_sample_residuals_by_bin_ instance-attribute

out_sample_residuals_by_bin_ = None

creation_date instance-attribute

creation_date = pd.Timestamp.today().strftime(
    "%Y-%m-%d %H:%M:%S"
)

is_fitted instance-attribute

is_fitted = False

fit_date instance-attribute

fit_date = None

skforecast_version instance-attribute

skforecast_version = __version__

python_version instance-attribute

python_version = sys.version.split(' ')[0]

forecaster_id instance-attribute

forecaster_id = forecaster_id

lags_are_contiguous instance-attribute

lags_are_contiguous = (
    self.lags is not None
    and np.array_equal(
        self.lags, np.arange(1, self.max_lag + 1)
    )
)

window_size instance-attribute

window_size = max(
    [
        ws
        for ws in [
            self.max_lag,
            self.max_size_window_features,
        ]
        if ws is not None
    ]
)

window_features_class_names instance-attribute

window_features_class_names = None

categorical_encoder instance-attribute

categorical_encoder = OrdinalEncoder(
    dtype=float,
    handle_unknown="use_encoded_value",
    unknown_value=np.nan,
    encoded_missing_value=np.nan,
).set_output(transform="pandas")

fit_kwargs instance-attribute

fit_kwargs = check_select_fit_kwargs(
    estimator=estimator, fit_kwargs=fit_kwargs
)

binner instance-attribute

binner = {}

binner_intervals_ instance-attribute

binner_intervals_ = {}

binner_kwargs instance-attribute

binner_kwargs = binner_kwargs

Methods:

create_train_X_y

create_train_X_y(
    series, exog=None, suppress_warnings=False
)

Create training matrices from multiple time series and exogenous variables. See Notes section for more details depending on the type of series and exog.

Parameters:

Name	Type	Description	Default
series	pandas DataFrame, dict	Training time series.	required
exog	pandas Series, pandas DataFrame, dict	Exogenous variable/s included as predictor/s.	None
suppress_warnings	bool	If True, skforecast warnings will be suppressed during the creation of the training matrices. See skforecast.exceptions.warn_skforecast_categories for more information.	False

Returns:

Name	Type	Description
X_train	pandas DataFrame	Training values (predictors).
y_train	pandas Series	Values (target) of the time series related to each row of X_train.

Notes

• If series is a wide-format pandas DataFrame, each column represents a different time series, and the index must be either a DatetimeIndex or a RangeIndex with frequency or step size, as appropriate
• If series is a long-format pandas DataFrame with a MultiIndex, the first level of the index must contain the series IDs, and the second level must be a DatetimeIndex with the same frequency across all series.
• If series is a dictionary, each key must be a series ID, and each value must be a named pandas Series. All series must have the same index, which must be either a DatetimeIndex or a RangeIndex, and they must share the same frequency or step size, as appropriate.
• If exog is a wide-format pandas DataFrame, it must share the same index type as series. Each column represents a different exogenous variable, and the same values are applied to all time series.
• If exog is a long-format pandas Series or DataFrame with a MultiIndex, the first level contains the series IDs to which it belongs, and the second level must be a pandas DatetimeIndex. Each exogenous variable must be represented as a separate column.
• If exog is a dictionary, each key must correspond to a series ID, and each value must be either a named pandas Series or DataFrame with the same index type as series, or None. It is not required for all series to contain all exogenous variables, but data types must be consistent across series for each variable.

Source code in skforecast/recursive/_forecaster_recursive_multiseries.py

@manage_warnings
def create_train_X_y(
    self,
    series: pd.DataFrame | dict[str, pd.Series | pd.DataFrame],
    exog: pd.Series | pd.DataFrame | dict[str, pd.Series | pd.DataFrame] | None = None,
    suppress_warnings: bool = False
) -> tuple[pd.DataFrame, pd.Series]:
    """
    Create training matrices from multiple time series and exogenous
    variables. See Notes section for more details depending on the type of
    `series` and `exog`.

    Parameters
    ----------
    series : pandas DataFrame, dict
        Training time series.
    exog : pandas Series, pandas DataFrame, dict, default None
        Exogenous variable/s included as predictor/s.
    suppress_warnings : bool, default False
        If `True`, skforecast warnings will be suppressed during the creation
        of the training matrices. See skforecast.exceptions.warn_skforecast_categories 
        for more information.

    Returns
    -------
    X_train : pandas DataFrame
        Training values (predictors).
    y_train : pandas Series
        Values (target) of the time series related to each row of `X_train`.

    Notes
    -----
    - If `series` is a wide-format pandas DataFrame, each column represents a
    different time series, and the index must be either a `DatetimeIndex` or 
    a `RangeIndex` with frequency or step size, as appropriate
    - If `series` is a long-format pandas DataFrame with a MultiIndex, the 
    first level of the index must contain the series IDs, and the second 
    level must be a `DatetimeIndex` with the same frequency across all series.
    - If series is a dictionary, each key must be a series ID, and each value 
    must be a named pandas Series. All series must have the same index, which 
    must be either a `DatetimeIndex` or a `RangeIndex`, and they must share the 
    same frequency or step size, as appropriate.
    - If `exog` is a wide-format pandas DataFrame, it must share the same 
    index type as series. Each column represents a different exogenous variable, 
    and the same values are applied to all time series.
    - If `exog` is a long-format pandas Series or DataFrame with a MultiIndex, 
    the first level contains the series IDs to which it belongs, and the second 
    level must be a pandas `DatetimeIndex`. Each exogenous variable must be 
    represented as a separate column.
    - If `exog` is a dictionary, each key must correspond to a series ID, and 
    each value must be either a named pandas `Series` or `DataFrame` with the 
    same index type as `series`, or `None`. It is not required for all series 
    to contain all exogenous variables, but data types must be consistent 
    across series for each variable.

    """

    output = self._create_train_X_y(
                 series            = series, 
                 exog              = exog, 
                 store_last_window = False
             )

    X_train = output[0]
    y_train = output[1]

    if self.encoding is None:
        X_train = X_train.drop(columns='_level_skforecast')

    return X_train, y_train

create_sample_weights

create_sample_weights(series_names_in_, X_train)

Create weights for each observation according to the forecaster's attributes series_weights and weight_func. The resulting weights are product of both types of weights.

Parameters:

Name	Type	Description	Default
series_names_in_	list	Names of the series (levels) used during training.	required
X_train	pandas DataFrame	Dataframe created with the create_train_X_y method, first return.	required

Returns:

Name	Type	Description
weights	numpy ndarray	Weights to use in fit method.

Source code in skforecast/recursive/_forecaster_recursive_multiseries.py

def create_sample_weights(
    self,
    series_names_in_: list,
    X_train: pd.DataFrame
) -> np.ndarray:
    """
    Create weights for each observation according to the forecaster's attributes
    `series_weights` and `weight_func`. The resulting weights are product of both
    types of weights.

    Parameters
    ----------
    series_names_in_ : list
        Names of the series (levels) used during training.
    X_train : pandas DataFrame
        Dataframe created with the `create_train_X_y` method, first return.

    Returns
    -------
    weights : numpy ndarray
        Weights to use in `fit` method.

    """

    weights = None
    weights_samples = None
    series_weights = None

    if self.series_weights is not None:
        # Series not present in series_weights have a weight of 1 in all their samples.
        # Keys in series_weights not present in series are ignored.
        series_not_in_series_weights = (
            set(series_names_in_) - set(self.series_weights.keys())
        )
        if series_not_in_series_weights:
            warnings.warn(
                f"{series_not_in_series_weights} not present in `series_weights`. "
                f"A weight of 1 is given to all their samples.",
                IgnoredArgumentWarning
            )
        self.series_weights_ = {col: 1. for col in series_names_in_}
        self.series_weights_.update(
            {
                k: v
                for k, v in self.series_weights.items()
                if k in self.series_weights_
            }
        )

        if self.encoding == "onehot":
            series_weights = [
                np.repeat(self.series_weights_[serie], sum(X_train[serie]))
                for serie in series_names_in_
            ]
        else:
            series_weights = [
                np.repeat(
                    self.series_weights_[serie],
                    sum(X_train["_level_skforecast"] == self.encoding_mapping_[serie]),
                )
                for serie in series_names_in_
            ]

        series_weights = np.concatenate(series_weights)

    if self.weight_func is not None:
        if isinstance(self.weight_func, Callable):
            self.weight_func_ = {
                col: copy(self.weight_func) for col in series_names_in_
            }
        else:
            # Series not present in weight_func have a weight of 1 in all their samples
            series_not_in_weight_func = (
                set(series_names_in_) - set(self.weight_func.keys())
            )
            if series_not_in_weight_func:
                warnings.warn(
                    f"{series_not_in_weight_func} not present in `weight_func`. "
                    f"A weight of 1 is given to all their samples.",
                    IgnoredArgumentWarning
                )
            self.weight_func_ = {
                col: self._weight_func_all_1 for col in series_names_in_
            }
            self.weight_func_.update(
                {
                    k: v
                    for k, v in self.weight_func.items()
                    if k in self.weight_func_
                }
            )

        weights_samples = []
        for key in self.weight_func_.keys():
            if self.encoding == "onehot":
                idx = X_train.index[X_train[key] == 1.0]
            else:
                idx = X_train.index[
                    X_train["_level_skforecast"] == self.encoding_mapping_[key]
                ]
            weights_samples.append(self.weight_func_[key](idx))
        weights_samples = np.concatenate(weights_samples)

    if series_weights is not None:
        weights = series_weights
        if weights_samples is not None:
            weights = weights * weights_samples
    else:
        if weights_samples is not None:
            weights = weights_samples

    if weights is not None:
        if np.isnan(weights).any():
            raise ValueError(
                "The resulting `weights` cannot have NaN values."
            )
        if np.any(weights < 0):
            raise ValueError(
                "The resulting `weights` cannot have negative values."
            )
        if np.sum(weights) == 0:
            raise ValueError(
                "The resulting `weights` cannot be normalized because "
                "the sum of the weights is zero."
            )

    return weights

fit

fit(
    series,
    exog=None,
    store_last_window=True,
    store_in_sample_residuals=False,
    random_state=123,
    suppress_warnings=False,
)

Training Forecaster. See Notes section for more details depending on the type of series and exog.

Additional arguments to be passed to the fit method of the estimator can be added with the fit_kwargs argument when initializing the forecaster.

Parameters:

Name	Type	Description	Default
series	pandas DataFrame, dict	Training time series.	required
exog	pandas Series, pandas DataFrame, dict	Exogenous variable/s included as predictor/s.	None
store_last_window	(bool, list)	Whether or not to store the last window (last_window_) of training data. If True, last window is stored for all series. If list, last window is stored for the series present in the list. If False, last window is not stored.	True
store_in_sample_residuals	bool	If True, in-sample residuals will be stored in the forecaster object after fitting (in_sample_residuals_ and in_sample_residuals_by_bin_ attributes). If False, only the intervals of the bins are stored.	False
random_state	int	Set a seed for the random generator so that the stored sample residuals are always deterministic.	123
suppress_warnings	bool	If True, skforecast warnings will be suppressed during the training process. See skforecast.exceptions.warn_skforecast_categories for more information.	False

Returns:

Type	Description
None

Notes

• If series is a wide-format pandas DataFrame, each column represents a different time series, and the index must be either a DatetimeIndex or a RangeIndex with frequency or step size, as appropriate
• If series is a long-format pandas DataFrame with a MultiIndex, the first level of the index must contain the series IDs, and the second level must be a DatetimeIndex with the same frequency across all series.
• If series is a dictionary, each key must be a series ID, and each value must be a named pandas Series. All series must have the same index, which must be either a DatetimeIndex or a RangeIndex, and they must share the same frequency or step size, as appropriate.
• If exog is a wide-format pandas DataFrame, it must share the same index type as series. Each column represents a different exogenous variable, and the same values are applied to all time series.
• If exog is a long-format pandas Series or DataFrame with a MultiIndex, the first level contains the series IDs to which it belongs, and the second level must be a pandas DatetimeIndex. Each exogenous variable must be represented as a separate column.
• If exog is a dictionary, each key must correspond to a series ID, and each value must be either a named pandas Series or DataFrame with the same index type as series, or None. It is not required for all series to contain all exogenous variables, but data types must be consistent across series for each variable.

Source code in skforecast/recursive/_forecaster_recursive_multiseries.py

@manage_warnings
def fit(
    self,
    series: pd.DataFrame | dict[str, pd.Series | pd.DataFrame],
    exog: pd.Series | pd.DataFrame | dict[str, pd.Series | pd.DataFrame] | None = None,
    store_last_window: bool | list[str] = True,
    store_in_sample_residuals: bool = False,
    random_state: int = 123,
    suppress_warnings: bool = False
) -> None:
    """
    Training Forecaster. See Notes section for more details depending on 
    the type of `series` and `exog`.

    Additional arguments to be passed to the `fit` method of the estimator 
    can be added with the `fit_kwargs` argument when initializing the forecaster.

    Parameters
    ----------
    series : pandas DataFrame, dict
        Training time series.
    exog : pandas Series, pandas DataFrame, dict, default None
        Exogenous variable/s included as predictor/s.
    store_last_window : bool, list, default True
        Whether or not to store the last window (`last_window_`) of training data.

        - If `True`, last window is stored for all series. 
        - If `list`, last window is stored for the series present in the list.
        - If `False`, last window is not stored.
    store_in_sample_residuals : bool, default False
        If `True`, in-sample residuals will be stored in the forecaster object
        after fitting (`in_sample_residuals_` and `in_sample_residuals_by_bin_`
        attributes).
        If `False`, only the intervals of the bins are stored.
    random_state : int, default 123
        Set a seed for the random generator so that the stored sample 
        residuals are always deterministic.
    suppress_warnings : bool, default False
        If `True`, skforecast warnings will be suppressed during the training 
        process. See skforecast.exceptions.warn_skforecast_categories for more
        information.

    Returns
    -------
    None

    Notes
    -----
    - If `series` is a wide-format pandas DataFrame, each column represents a
    different time series, and the index must be either a `DatetimeIndex` or 
    a `RangeIndex` with frequency or step size, as appropriate
    - If `series` is a long-format pandas DataFrame with a MultiIndex, the 
    first level of the index must contain the series IDs, and the second 
    level must be a `DatetimeIndex` with the same frequency across all series.
    - If series is a dictionary, each key must be a series ID, and each value 
    must be a named pandas Series. All series must have the same index, which 
    must be either a `DatetimeIndex` or a `RangeIndex`, and they must share the 
    same frequency or step size, as appropriate.
    - If `exog` is a wide-format pandas DataFrame, it must share the same 
    index type as series. Each column represents a different exogenous variable, 
    and the same values are applied to all time series.
    - If `exog` is a long-format pandas Series or DataFrame with a MultiIndex, 
    the first level contains the series IDs to which it belongs, and the second 
    level must be a pandas `DatetimeIndex`. Each exogenous variable must be 
    represented as a separate column.
    - If `exog` is a dictionary, each key must correspond to a series ID, and 
    each value must be either a named pandas `Series` or `DataFrame` with the 
    same index type as `series`, or `None`. It is not required for all series 
    to contain all exogenous variables, but data types must be consistent 
    across series for each variable.

    """

    # TODO: create a method reset_forecaster() to reset all attributes
    # Reset values in case the forecaster has already been fitted.
    self.last_window_                         = None
    self.index_type_                          = None
    self.index_freq_                          = None
    self.training_range_                      = None
    self.series_names_in_                     = None
    self.exog_in_                             = False
    self.exog_names_in_                       = None
    self.exog_type_in_                        = None
    self.exog_dtypes_in_                      = None
    self.exog_dtypes_out_                     = None
    self.categorical_features_names_in_       = None
    self.X_train_series_names_in_             = None
    self.X_train_window_features_names_out_   = None
    self.X_train_calendar_features_names_out_ = None
    self.X_train_exog_names_out_              = None
    self.X_train_features_names_out_          = None
    self.encoding_mapping_                    = {}
    self.in_sample_residuals_                 = None
    self.in_sample_residuals_by_bin_          = None
    self.out_sample_residuals_                = None
    self.out_sample_residuals_by_bin_         = None
    self.binner                               = {}
    self.binner_intervals_                    = {}
    self.is_fitted                            = False
    self.fit_date                             = None

    (
        X_train,
        y_train,
        series_indexes,
        series_names_in_,
        X_train_series_names_in_,
        exog_names_in_,
        categorical_features_names_in_,
        X_train_window_features_names_out_,
        X_train_calendar_features_names_out_,
        X_train_exog_names_out_,
        exog_dtypes_in_,
        exog_dtypes_out_,
        last_window_
    ) = self._create_train_X_y(
            series=series, exog=exog, store_last_window=store_last_window
        )

    sample_weight = self.create_sample_weights(
                        series_names_in_ = series_names_in_,
                        X_train          = X_train
                    )

    X_train_estimator = (
        X_train
        if self.encoding is not None
        else X_train.drop(columns="_level_skforecast")
    )
    X_train_features_names_out_ = X_train_estimator.columns.to_list()

    if self.categorical_features is not None:
        all_categorical_names = (
            list(categorical_features_names_in_) if categorical_features_names_in_ else []
        )
        if self.encoding == 'ordinal_category':
            all_categorical_names.append('_level_skforecast')
        fit_kwargs = configure_estimator_categorical_features(
                         estimator                      = self.estimator,
                         categorical_features_names_in_ = all_categorical_names,
                         X_train_features_names_out_    = X_train_features_names_out_,
                         fit_kwargs                     = {**self.fit_kwargs}
                     )
    else:
        fit_kwargs = {**self.fit_kwargs}

    X_train_estimator = cast_catboost_categorical_columns_dataframe(
        X              = X_train_estimator,
        fit_kwargs     = fit_kwargs,
        estimator      = self.estimator,
        feature_names  = X_train_features_names_out_,
    )

    if sample_weight is not None:
        self.estimator.fit(
            X             = X_train_estimator,
            y             = y_train,
            sample_weight = sample_weight,
            **fit_kwargs
        )
    else:
        self.estimator.fit(X=X_train_estimator, y=y_train, **fit_kwargs)

    self.series_names_in_ = series_names_in_
    self.X_train_series_names_in_ = X_train_series_names_in_
    self.X_train_window_features_names_out_ = X_train_window_features_names_out_
    self.X_train_calendar_features_names_out_ = X_train_calendar_features_names_out_
    self.X_train_features_names_out_ = X_train_features_names_out_

    self.is_fitted = True
    self.fit_date = pd.Timestamp.today().strftime('%Y-%m-%d %H:%M:%S')
    self.training_range_ = {k: v[[0, -1]] for k, v in series_indexes.items()}
    self.index_type_ = type(series_indexes[series_names_in_[0]])
    if isinstance(series_indexes[series_names_in_[0]], pd.DatetimeIndex):
        self.index_freq_ = series_indexes[series_names_in_[0]].freq
    else:
        self.index_freq_ = series_indexes[series_names_in_[0]].step

    # NOTE: When `exog` doesn't match series IDs, exogs are not included in the
    # training matrices, X_train_exog_names_out_ is None and the forecaster 
    # will be considered trained without exogenous variables.
    if exog is not None and X_train_exog_names_out_ is not None:
        self.exog_in_ = True
        self.exog_names_in_ = exog_names_in_
        self.exog_type_in_ = type(exog)
        self.exog_dtypes_in_ = exog_dtypes_in_
        self.exog_dtypes_out_ = exog_dtypes_out_
        self.categorical_features_names_in_ = categorical_features_names_in_
        self.X_train_exog_names_out_ = X_train_exog_names_out_

    self.in_sample_residuals_ = {}
    self.in_sample_residuals_by_bin_ = {}
    if self._probabilistic_mode is not False:
        y_train = y_train.to_numpy()
        y_pred = self.estimator.predict(X_train_estimator)
        if self.encoding is not None:
            for level in X_train_series_names_in_:
                if self.encoding == 'onehot':
                    mask = X_train[level].to_numpy() == 1.
                else:
                    encoded_value = self.encoding_mapping_[level]
                    mask = X_train['_level_skforecast'].to_numpy() == encoded_value

                self._binning_in_sample_residuals(
                    level                     = level,
                    y_true                    = y_train[mask],
                    y_pred                    = y_pred[mask],
                    store_in_sample_residuals = store_in_sample_residuals,
                    random_state              = random_state
                )

        # NOTE: the _unknown_level is a random sample of 10_000 residuals of all levels.
        self._binning_in_sample_residuals(
            level                     = '_unknown_level',
            y_true                    = y_train,
            y_pred                    = y_pred,
            store_in_sample_residuals = store_in_sample_residuals,
            random_state              = random_state
        )

    if not store_in_sample_residuals:
        # NOTE: create empty dictionaries to avoid errors when calling predict()
        if self.encoding is not None:
            for level in X_train_series_names_in_:
                self.in_sample_residuals_[level] = None
                self.in_sample_residuals_by_bin_[level] = None
        self.in_sample_residuals_['_unknown_level'] = None
        self.in_sample_residuals_by_bin_['_unknown_level'] = None

    if store_last_window:
        self.last_window_ = last_window_

create_predict_X

create_predict_X(
    steps,
    levels=None,
    last_window=None,
    exog=None,
    suppress_warnings=False,
    check_inputs=True,
)

Create the predictors needed to predict steps ahead. As it is a recursive process, the predictors are created at each iteration of the prediction process.

Parameters:

Name	Type	Description	Default
steps	int	Number of steps to predict.	required
levels	(str, list)	Time series to be predicted. If None all levels whose last window ends at the same datetime index will be predicted together.	None
last_window	pandas DataFrame	Series values used to create the predictors (lags) needed in the first iteration of the prediction (t + 1). If last_window = None, the values stored in self.last_window_ are used to calculate the initial predictors, and the predictions start right after training data.	None
exog	pandas Series, pandas DataFrame	Exogenous variable/s included as predictor/s.	None
suppress_warnings	bool	If True, skforecast warnings will be suppressed during the prediction process. See skforecast.exceptions.warn_skforecast_categories for more information.	False
check_inputs	bool	If True, the input is checked for possible warnings and errors with the check_predict_input function. This argument is created for internal use and is not recommended to be changed.	True

Returns:

Name	Type	Description
X_predict	pandas DataFrame	Long-format DataFrame with the predictors. The columns are level and one column for each predictor. The index is the same as the prediction index.

Source code in skforecast/recursive/_forecaster_recursive_multiseries.py

@manage_warnings
def create_predict_X(
    self,
    steps: int,
    levels: str | list[str] | None = None,
    last_window: pd.DataFrame | None = None,
    exog: pd.Series | pd.DataFrame | dict[str, pd.Series | pd.DataFrame] | None = None,
    suppress_warnings: bool = False,
    check_inputs: bool = True
) -> pd.DataFrame:
    """
    Create the predictors needed to predict `steps` ahead. As it is a recursive
    process, the predictors are created at each iteration of the prediction 
    process.

    Parameters
    ----------
    steps : int
        Number of steps to predict. 
    levels : str, list, default None
        Time series to be predicted. If `None` all levels whose last window
        ends at the same datetime index will be predicted together.
    last_window : pandas DataFrame, default None
        Series values used to create the predictors (lags) needed in the 
        first iteration of the prediction (t + 1).
        If `last_window = None`, the values stored in `self.last_window_` are
        used to calculate the initial predictors, and the predictions start
        right after training data.
    exog : pandas Series, pandas DataFrame, default None
        Exogenous variable/s included as predictor/s.
    suppress_warnings : bool, default False
        If `True`, skforecast warnings will be suppressed during the prediction 
        process. See skforecast.exceptions.warn_skforecast_categories for more
        information.
    check_inputs : bool, default True
        If `True`, the input is checked for possible warnings and errors 
        with the `check_predict_input` function. This argument is created 
        for internal use and is not recommended to be changed.

    Returns
    -------
    X_predict : pandas DataFrame
        Long-format DataFrame with the predictors. The columns are `level` and 
        one column for each predictor. The index is the same as the prediction 
        index.

    """

    (
        last_window,
        exog_values_dict,
        calendar_values,
        levels,
        prediction_index,
        _
    ) = self._create_predict_inputs(
            steps        = steps,
            levels       = levels,
            last_window  = last_window,
            exog         = exog,
            check_inputs = check_inputs
        )

    with warnings.catch_warnings():
        warnings.filterwarnings(
            "ignore", 
            message="X does not have valid feature names", 
            category=UserWarning
        )
        predictions = self._recursive_predict(
                          steps            = steps,
                          levels           = levels,
                          last_window      = last_window,
                          exog_values_dict = exog_values_dict,
                          calendar_values  = calendar_values
                      )

    if self.lags is not None:
        idx_lags = np.arange(-steps, 0)[:, None] - self.lags
    len_X_train_series_names_in_ = len(self.X_train_series_names_in_)
    exog_shape = len(self.X_train_exog_names_out_) if exog is not None else 0

    X_predict = []
    for i, level in enumerate(levels):

        X_predict_level = []
        full_predictors_level = np.concatenate(
            (last_window[level].to_numpy(), predictions[:, i])
        )

        if self.lags is not None:
            X_predict_level.append(
                full_predictors_level[idx_lags + len(full_predictors_level)]
            )

        if self.window_features is not None:
            X_window_features = np.full(
                shape      = (steps, len(self.X_train_window_features_names_out_)), 
                fill_value = np.nan, 
                order      = 'C',
                dtype      = float
            )
            for j in range(steps):
                X_window_features[j, :] = np.concatenate(
                    [
                        wf.transform(full_predictors_level[j:-(steps - j)]) 
                        for wf in self.window_features
                    ]
                )
            X_predict_level.append(X_window_features)

        if self.encoding is not None:
            if self.encoding == 'onehot':
                level_encoded = np.zeros(
                                    shape = (1, len_X_train_series_names_in_),
                                    dtype = float
                                )
                level_encoded[0][self.X_train_series_names_in_.index(level)] = 1.
            else:
                level_encoded = np.array(
                                    [self.encoding_mapping_.get(level, None)],
                                    dtype = 'float64'
                                )

            level_encoded = np.tile(level_encoded, (steps, 1))
            X_predict_level.append(level_encoded)

        if exog is not None:
            exog_cols = np.full(
                shape=(steps, exog_shape), fill_value=np.nan, order='C', dtype=float
            )
            for j in range(steps):
                exog_cols[j, :] = exog_values_dict[j + 1][i, :]
            X_predict_level.append(exog_cols)

        if calendar_values is not None:
            # Calendar features are shared by all series (levels).
            X_predict_level.append(calendar_values)

        X_predict.append(np.concatenate(X_predict_level, axis=1))

    X_predict = pd.DataFrame(
                    data    = (
                        np.concatenate(X_predict) if levels
                        else np.empty((0, len(self.X_train_features_names_out_)))
                    ),
                    index   = np.tile(prediction_index, len(levels)),
                    columns = self.X_train_features_names_out_
                )
    X_predict.insert(0, 'level', np.repeat(levels, steps))

    # NOTE: Order needed to have the same structure as the output of predict methods.
    order_dict = {level: i for i, level in enumerate(levels)}
    X_predict['order'] = X_predict['level'].map(order_dict)
    X_predict = (
        X_predict
        .reset_index()
        .sort_values(by=['index', 'order'])
        .set_index('index')
        .rename_axis(index=None)
        .drop(columns='order')
    )

    if self.exog_in_:
        categorical_features = any(
            not pd.api.types.is_numeric_dtype(dtype) or pd.api.types.is_bool_dtype(dtype) 
            for dtype in set(self.exog_dtypes_out_.values())
        )
        if categorical_features:
            X_predict = X_predict.astype(self.exog_dtypes_out_, copy=False)

    if self.transformer_series is not None or self.differentiation is not None:
        warnings.warn(
            "The output matrix is in the transformed scale due to the "
            "inclusion of transformations or differentiation in the Forecaster. "
            "As a result, any predictions generated using this matrix will also "
            "be in the transformed scale. Please refer to the documentation "
            "for more details: "
            "https://skforecast.org/latest/user_guides/training-and-prediction-matrices.html",
            DataTransformationWarning
        )

    return X_predict

predict

predict(
    steps,
    levels=None,
    last_window=None,
    exog=None,
    suppress_warnings=False,
    check_inputs=True,
)

Predict n steps ahead. It is a recursive process in which, each prediction, is used as a predictor for the next step. Only levels whose last window ends at the same datetime index can be predicted together.

Parameters:

Name	Type	Description	Default
steps	int	Number of steps to predict.	required
levels	(str, list)	Time series to be predicted. If None all levels whose last window ends at the same datetime index will be predicted together.	None
last_window	pandas DataFrame	Series values used to create the predictors (lags) needed in the first iteration of the prediction (t + 1). If last_window = None, the values stored in self.last_window_ are used to calculate the initial predictors, and the predictions start right after training data.	None
exog	pandas Series, pandas DataFrame, dict	Exogenous variable/s included as predictor/s.	None
suppress_warnings	bool	If True, skforecast warnings will be suppressed during the prediction process. See skforecast.exceptions.warn_skforecast_categories for more information.	False
check_inputs	bool	If True, the input is checked for possible warnings and errors with the check_predict_input function. This argument is created for internal use and is not recommended to be changed.	True

Returns:

Name	Type	Description
predictions	pandas DataFrame	Long-format DataFrame with the predictions. The columns are level and pred.

Source code in skforecast/recursive/_forecaster_recursive_multiseries.py

@manage_warnings
def predict(
    self,
    steps: int,
    levels: str | list[str] | None = None,
    last_window: pd.DataFrame | None = None,
    exog: pd.Series | pd.DataFrame | dict[str, pd.Series | pd.DataFrame] | None = None,
    suppress_warnings: bool = False,
    check_inputs: bool = True
) -> pd.DataFrame:
    """
    Predict n steps ahead. It is a recursive process in which, each prediction,
    is used as a predictor for the next step. Only levels whose last window
    ends at the same datetime index can be predicted together.

    Parameters
    ----------
    steps : int
        Number of steps to predict. 
    levels : str, list, default None
        Time series to be predicted. If `None` all levels whose last window
        ends at the same datetime index will be predicted together.
    last_window : pandas DataFrame, default None
        Series values used to create the predictors (lags) needed in the 
        first iteration of the prediction (t + 1).
        If `last_window = None`, the values stored in `self.last_window_` are
        used to calculate the initial predictors, and the predictions start
        right after training data.
    exog : pandas Series, pandas DataFrame, dict, default None
        Exogenous variable/s included as predictor/s.
    suppress_warnings : bool, default False
        If `True`, skforecast warnings will be suppressed during the prediction 
        process. See skforecast.exceptions.warn_skforecast_categories for more
        information.
    check_inputs : bool, default True
        If `True`, the input is checked for possible warnings and errors 
        with the `check_predict_input` function. This argument is created 
        for internal use and is not recommended to be changed.

    Returns
    -------
    predictions : pandas DataFrame
        Long-format DataFrame with the predictions. The columns are `level`
        and `pred`.

    """

    (
        last_window,
        exog_values_dict,
        calendar_values,
        levels,
        prediction_index,
        differentiators
    ) = self._create_predict_inputs(
            steps        = steps,
            levels       = levels,
            last_window  = last_window,
            exog         = exog,
            check_inputs = check_inputs
        )

    with warnings.catch_warnings():
        warnings.filterwarnings(
            "ignore", 
            message="X does not have valid feature names", 
            category=UserWarning
        )
        predictions = self._recursive_predict(
                          steps            = steps,
                          levels           = levels,
                          last_window      = last_window,
                          exog_values_dict = exog_values_dict,
                          calendar_values  = calendar_values
                      )

    for i, level in enumerate(levels):
        if differentiators.get(level) is not None:
            predictions[:, i] = (
                differentiators[level]
                .inverse_transform_next_window(predictions[:, i])
            )

        predictions[:, i] = transform_numpy(
            array             = predictions[:, i],
            transformer       = self.transformer_series_.get(level, self.transformer_series_['_unknown_level']),
            fit               = False,
            inverse_transform = True
        )

    n_steps, n_levels = predictions.shape
    predictions = pd.DataFrame(
        {"level": np.tile(levels, n_steps), "pred": predictions.ravel()},
        index = np.repeat(prediction_index, n_levels),
    )

    return predictions

predict_bootstrapping

predict_bootstrapping(
    steps,
    levels=None,
    last_window=None,
    exog=None,
    n_boot=250,
    use_in_sample_residuals=True,
    use_binned_residuals=True,
    random_state=123,
    suppress_warnings=False,
)

Generate multiple forecasting predictions using a bootstrapping process. By sampling from a collection of past observed errors (the residuals), each iteration of bootstrapping generates a different set of predictions. Only levels whose last window ends at the same datetime index can be predicted together. See the References section for more information.

Parameters:

Name	Type	Description	Default
steps	int	Number of steps to predict.	required
levels	(str, list)	Time series to be predicted. If None all levels whose last window ends at the same datetime index will be predicted together.	None
last_window	pandas DataFrame	Series values used to create the predictors (lags) needed in the first iteration of the prediction (t + 1). If last_window = None, the values stored in self.last_window_ are used to calculate the initial predictors, and the predictions start right after training data.	None
exog	pandas Series, pandas DataFrame, dict	Exogenous variable/s included as predictor/s.	None
n_boot	int	Number of bootstrapping iterations to perform when estimating prediction intervals.	250
use_in_sample_residuals	bool	If True, residuals from the training data are used as proxy of prediction error to create predictions. If False, out of sample residuals (calibration) are used. Out-of-sample residuals must be precomputed using Forecaster's set_out_sample_residuals() method.	True
use_binned_residuals	bool	If True, residuals are selected based on the predicted values (binned selection). If False, residuals are selected randomly. New in version 0.15.0	True
random_state	int	Seed for the random number generator to ensure reproducibility.	123
suppress_warnings	bool	If True, skforecast warnings will be suppressed during the prediction process. See skforecast.exceptions.warn_skforecast_categories for more information.	False

Returns:

Name	Type	Description
boot_predictions	pandas DataFrame	Long-format DataFrame with the bootstrapping predictions. The columns are level, pred_boot_0, pred_boot_1, ..., pred_boot_n_boot.

References

.. [1] Forecasting: Principles and Practice (3^rd ed) Rob J Hyndman and George Athanasopoulos. https://otexts.com/fpp3/prediction-intervals.html

Source code in skforecast/recursive/_forecaster_recursive_multiseries.py

@manage_warnings
def predict_bootstrapping(
    self,
    steps: int,
    levels: str | list[str] | None = None,
    last_window: pd.DataFrame | None = None,
    exog: pd.Series | pd.DataFrame | dict[str, pd.Series | pd.DataFrame] | None = None,
    n_boot: int = 250,
    use_in_sample_residuals: bool = True,
    use_binned_residuals: bool = True,
    random_state: int = 123,
    suppress_warnings: bool = False
) -> pd.DataFrame:
    """
    Generate multiple forecasting predictions using a bootstrapping process. 
    By sampling from a collection of past observed errors (the residuals),
    each iteration of bootstrapping generates a different set of predictions. 
    Only levels whose last window ends at the same datetime index can be 
    predicted together. See the References section for more information. 

    Parameters
    ----------
    steps : int
        Number of steps to predict. 
    levels : str, list, default None
        Time series to be predicted. If `None` all levels whose last window
        ends at the same datetime index will be predicted together.
    last_window : pandas DataFrame, default None
        Series values used to create the predictors (lags) needed in the 
        first iteration of the prediction (t + 1).
        If `last_window = None`, the values stored in `self.last_window_` are
        used to calculate the initial predictors, and the predictions start
        right after training data.
    exog : pandas Series, pandas DataFrame, dict, default None
        Exogenous variable/s included as predictor/s.
    n_boot : int, default 250
        Number of bootstrapping iterations to perform when estimating prediction
        intervals.
    use_in_sample_residuals : bool, default True
        If `True`, residuals from the training data are used as proxy of
        prediction error to create predictions. 
        If `False`, out of sample residuals (calibration) are used. 
        Out-of-sample residuals must be precomputed using Forecaster's
        `set_out_sample_residuals()` method.
    use_binned_residuals : bool, default True
        If `True`, residuals are selected based on the predicted values 
        (binned selection).
        If `False`, residuals are selected randomly.
        **New in version 0.15.0**
    random_state : int, default 123
        Seed for the random number generator to ensure reproducibility.
    suppress_warnings : bool, default False
        If `True`, skforecast warnings will be suppressed during the prediction 
        process. See skforecast.exceptions.warn_skforecast_categories for more
        information.

    Returns
    -------
    boot_predictions : pandas DataFrame
        Long-format DataFrame with the bootstrapping predictions. The columns
        are `level`, `pred_boot_0`, `pred_boot_1`, ..., `pred_boot_n_boot`.

    References
    ----------
    .. [1] Forecasting: Principles and Practice (3rd ed) Rob J Hyndman and George Athanasopoulos.
           https://otexts.com/fpp3/prediction-intervals.html

    """

    (
        last_window,
        exog_values_dict,
        calendar_values,
        levels,
        prediction_index,
        differentiators
    ) = self._create_predict_inputs(
            steps                   = steps,
            levels                  = levels,
            last_window             = last_window,
            exog                    = exog,
            predict_probabilistic   = True,
            use_in_sample_residuals = use_in_sample_residuals,
            use_binned_residuals    = use_binned_residuals
        )

    if use_in_sample_residuals:
        residuals = self.in_sample_residuals_
        residuals_by_bin = self.in_sample_residuals_by_bin_
    else:
        residuals = self.out_sample_residuals_
        residuals_by_bin = self.out_sample_residuals_by_bin_

    n_levels = len(levels)
    rng = np.random.default_rng(seed=random_state)

    if use_binned_residuals:
        # Pre-allocate 4D array directly: (n_bins, steps, n_boot, n_levels)
        # Loop order must match original to preserve RNG sequence for reproducibility
        level_binners = [
            self.binner.get(level, self.binner['_unknown_level'])
            for level in levels
        ]
        # Each level has its own binner and may have a different number of
        # bins, the array is sized for the level with the most bins.
        n_bins = max(binner.n_bins_ for binner in level_binners)
        sampled_residuals = np.empty(
            (n_bins, steps, n_boot, n_levels), order='C', dtype=float
        )
        for bin_idx in range(n_bins):
            for i, level in enumerate(levels):
                # Padding for levels with fewer bins, never indexed: NaN instead of
                # uninitialized memory so that a bad lookup surfaces as NaN.
                if bin_idx >= level_binners[i].n_bins_:
                    sampled_residuals[bin_idx, :, :, i] = np.nan
                    continue
                sampled_residuals[bin_idx, :, :, i] = rng.choice(
                    a       = residuals_by_bin.get(level, residuals_by_bin['_unknown_level'])[bin_idx],
                    size    = (steps, n_boot),
                    replace = True
                )
    else:
        sampled_residuals = np.full(
            shape      = (steps, n_levels, n_boot),
            fill_value = np.nan,
            order      = 'C',
            dtype      = float
        )
        for i, level in enumerate(levels):
            sampled_residuals[:, i, :] = rng.choice(
                a       = residuals.get(level, residuals['_unknown_level']),
                size    = (steps, n_boot),
                replace = True
            )

    boot_columns = [f"pred_boot_{i}" for i in range(n_boot)]

    with warnings.catch_warnings():
        warnings.filterwarnings(
            "ignore", 
            message="X does not have valid feature names", 
            category=UserWarning
        )
        boot_predictions = self._recursive_predict_bootstrapping(
            steps                = steps,
            levels               = levels,
            last_window          = last_window,
            n_boot               = n_boot,
            sampled_residuals    = sampled_residuals,
            use_binned_residuals = use_binned_residuals,
            exog_values_dict     = exog_values_dict,
            calendar_values      = calendar_values,
        )

    for i, level in enumerate(levels):

        if differentiators.get(level) is not None:
            boot_predictions[:, i, :] = (
                differentiators[level]
                .inverse_transform_next_window(boot_predictions[:, i, :])
            )

        transformer_level = self.transformer_series_.get(
                                level,
                                self.transformer_series_['_unknown_level']
                            )
        if transformer_level is not None:
            boot_predictions[:, i, :] = transform_numpy(
                array             = boot_predictions[:, i, :],
                transformer       = transformer_level,
                fit               = False,
                inverse_transform = True
            )

    boot_predictions = pd.DataFrame(
                           data    = boot_predictions.reshape(-1, n_boot),
                           index   = np.repeat(prediction_index, n_levels),
                           columns = boot_columns
                       )
    boot_predictions.insert(0, 'level', np.tile(levels, steps))

    return boot_predictions

predict_interval

predict_interval(
    steps,
    levels=None,
    last_window=None,
    exog=None,
    method="conformal",
    interval=[0.05, 0.95],
    n_boot=250,
    use_in_sample_residuals=True,
    use_binned_residuals=True,
    random_state=123,
    suppress_warnings=False,
)

Predict n steps ahead and estimate prediction intervals using either bootstrapping or conformal prediction methods. Refer to the References section for additional details on these methods.

Parameters:

Name	Type	Description	Default
steps	int	Number of steps to predict.	required
levels	(str, list)	Time series to be predicted. If None all levels whose last window ends at the same datetime index will be predicted together.	None
last_window	pandas DataFrame	Series values used to create the predictors (lags) needed in the first iteration of the prediction (t + 1). If last_window = None, the values stored in self.last_window_ are used to calculate the initial predictors, and the predictions start right after training data.	None
exog	pandas Series, pandas DataFrame, dict	Exogenous variable/s included as predictor/s.	None
method	str	Technique used to estimate prediction intervals. Available options: 'bootstrapping': Bootstrapping is used to generate prediction intervals [1]_. 'conformal': Employs the conformal prediction split method for interval estimation [2]_.	'conformal'
interval	(float, list, tuple)	Confidence level of the prediction interval. Interpretation depends on the method used: If float, represents the nominal (expected) coverage (between 0 and 1). For instance, interval=0.95 corresponds to [0.025, 0.975] quantiles. If list or tuple, defines the exact quantiles to compute, which must be between 0 and 1 inclusive. For example, interval of 95% should be as interval = [0.025, 0.975]. When using method='conformal', the interval must be a float or a list/tuple defining a symmetric interval. Changed in version 0.23.0: interval is now expressed as quantiles (0-1) instead of percentiles (0-100). Passing percentiles is not longer supported and will raise a ValueError.	[0.05, 0.95]
n_boot	int	Number of bootstrapping iterations to perform when estimating prediction intervals.	250
use_in_sample_residuals	bool	If True, residuals from the training data are used as proxy of prediction error to create predictions. If False, out of sample residuals (calibration) are used. Out-of-sample residuals must be precomputed using Forecaster's set_out_sample_residuals() method.	True
use_binned_residuals	bool	If True, residuals are selected based on the predicted values (binned selection). If False, residuals are selected randomly.	True
random_state	int	Seed for the random number generator to ensure reproducibility.	123
suppress_warnings	bool	If True, skforecast warnings will be suppressed during the prediction process. See skforecast.exceptions.warn_skforecast_categories for more information.	False

Returns:

Name	Type	Description
predictions	pandas DataFrame	Long-format DataFrame with the predictions and the lower and upper bounds of the estimated interval. The columns are level, pred, lower_bound, upper_bound.

References

.. [1] Forecasting: Principles and Practice (3^rd ed) Rob J Hyndman and George Athanasopoulos. https://otexts.com/fpp3/prediction-intervals.html

.. [2] MAPIE - Model Agnostic Prediction Interval Estimator. https://mapie.readthedocs.io/en/stable/content/conformal-prediction/regression/#2-the-split-method

Source code in skforecast/recursive/_forecaster_recursive_multiseries.py

@manage_warnings
def predict_interval(
    self,
    steps: int,
    levels: str | list[str] | None = None,
    last_window: pd.DataFrame | None = None,
    exog: pd.Series | pd.DataFrame | dict[str, pd.Series | pd.DataFrame] | None = None,
    method: str = 'conformal',
    interval: float | list[float] | tuple[float] = [0.05, 0.95],
    n_boot: int = 250,
    use_in_sample_residuals: bool = True,
    use_binned_residuals: bool = True,
    random_state: int = 123,
    suppress_warnings: bool = False
) -> pd.DataFrame:
    """
    Predict n steps ahead and estimate prediction intervals using either 
    bootstrapping or conformal prediction methods. Refer to the References 
    section for additional details on these methods.

    Parameters
    ----------
    steps : int
        Number of steps to predict. 
    levels : str, list, default None
        Time series to be predicted. If `None` all levels whose last window
        ends at the same datetime index will be predicted together.
    last_window : pandas DataFrame, default None
        Series values used to create the predictors (lags) needed in the 
        first iteration of the prediction (t + 1).
        If `last_window = None`, the values stored in `self.last_window_` are
        used to calculate the initial predictors, and the predictions start
        right after training data.
    exog : pandas Series, pandas DataFrame, dict, default None
        Exogenous variable/s included as predictor/s.
    method : str, default 'conformal'
        Technique used to estimate prediction intervals. Available options:

        - 'bootstrapping': Bootstrapping is used to generate prediction 
        intervals [1]_.
        - 'conformal': Employs the conformal prediction split method for 
        interval estimation [2]_.
    interval : float, list, tuple, default [0.05, 0.95]
        Confidence level of the prediction interval. Interpretation depends 
        on the method used:

        - If `float`, represents the nominal (expected) coverage (between 0 
        and 1). For instance, `interval=0.95` corresponds to `[0.025, 0.975]` 
        quantiles.
        - If `list` or `tuple`, defines the exact quantiles to compute, which 
        must be between 0 and 1 inclusive. For example, interval 
        of 95% should be as `interval = [0.025, 0.975]`.
        - When using `method='conformal'`, the interval must be a float or 
        a list/tuple defining a symmetric interval.

        **Changed in version 0.23.0:** `interval` is now expressed as
        quantiles (0-1) instead of percentiles (0-100). Passing percentiles
        is not longer supported and will raise a `ValueError`.
    n_boot : int, default 250
        Number of bootstrapping iterations to perform when estimating prediction
        intervals.
    use_in_sample_residuals : bool, default True
        If `True`, residuals from the training data are used as proxy of
        prediction error to create predictions. 
        If `False`, out of sample residuals (calibration) are used. 
        Out-of-sample residuals must be precomputed using Forecaster's
        `set_out_sample_residuals()` method.
    use_binned_residuals : bool, default True
        If `True`, residuals are selected based on the predicted values 
        (binned selection).
        If `False`, residuals are selected randomly.
    random_state : int, default 123
        Seed for the random number generator to ensure reproducibility.
    suppress_warnings : bool, default False
        If `True`, skforecast warnings will be suppressed during the prediction 
        process. See skforecast.exceptions.warn_skforecast_categories for more
        information.

    Returns
    -------
    predictions : pandas DataFrame
        Long-format DataFrame with the predictions and the lower and upper
        bounds of the estimated interval. The columns are `level`, `pred`,
        `lower_bound`, `upper_bound`.

    References
    ----------
    .. [1] Forecasting: Principles and Practice (3rd ed) Rob J Hyndman and George Athanasopoulos.
           https://otexts.com/fpp3/prediction-intervals.html

    .. [2] MAPIE - Model Agnostic Prediction Interval Estimator.
           https://mapie.readthedocs.io/en/stable/content/conformal-prediction/regression/#2-the-split-method

    """

    if method == "bootstrapping":

        if isinstance(interval, (list, tuple)):
            check_interval(interval=interval, ensure_symmetric_intervals=False)
            interval = np.array(interval)
        else:
            check_interval(alpha=interval, alpha_literal='interval')
            interval = np.array([0.5 - interval / 2, 0.5 + interval / 2])

        boot_predictions = self.predict_bootstrapping(
                               steps                   = steps,
                               levels                  = levels,
                               last_window             = last_window,
                               exog                    = exog,
                               n_boot                  = n_boot,
                               use_in_sample_residuals = use_in_sample_residuals,
                               use_binned_residuals    = use_binned_residuals,
                               random_state            = random_state,
                               suppress_warnings       = suppress_warnings
                           )

        predictions = self.predict(
                          steps             = steps,
                          levels            = levels,
                          last_window       = last_window,
                          exog              = exog,
                          suppress_warnings = suppress_warnings,
                          check_inputs      = False
                      )

        boot_predictions[['lower_bound', 'upper_bound']] = (
            boot_predictions.iloc[:, 1:].quantile(q=interval, axis=1).transpose()
        )
        predictions = pd.concat([
            predictions, boot_predictions[['lower_bound', 'upper_bound']]
        ], axis=1)

    elif method == 'conformal':

        if isinstance(interval, (list, tuple)):
            check_interval(interval=interval, ensure_symmetric_intervals=True)
            nominal_coverage = interval[1] - interval[0]
        else:
            check_interval(alpha=interval, alpha_literal='interval')
            nominal_coverage = interval

        predictions = self._predict_interval_conformal(
                          steps                   = steps,
                          levels                  = levels,
                          last_window             = last_window,
                          exog                    = exog,
                          nominal_coverage        = nominal_coverage,
                          use_in_sample_residuals = use_in_sample_residuals,
                          use_binned_residuals    = use_binned_residuals
                      )
    else:
        raise ValueError(
            f"Invalid `method` '{method}'. Choose 'bootstrapping' or 'conformal'."
        )

    return predictions

predict_quantiles

predict_quantiles(
    steps,
    levels=None,
    last_window=None,
    exog=None,
    quantiles=[0.05, 0.5, 0.95],
    n_boot=250,
    use_in_sample_residuals=True,
    use_binned_residuals=True,
    random_state=123,
    suppress_warnings=False,
)

Calculate the specified quantiles for each step. After generating multiple forecasting predictions through a bootstrapping process, each quantile is calculated for each step.

Parameters:

Name	Type	Description	Default
steps	int	Number of steps to predict.	required
levels	(str, list)	Time series to be predicted. If None all levels whose last window ends at the same datetime index will be predicted together.	None
last_window	pandas DataFrame	Series values used to create the predictors (lags) needed in the first iteration of the prediction (t + 1). If last_window = None, the values stored in self.last_window_ are used to calculate the initial predictors, and the predictions start right after training data.	None
exog	pandas Series, pandas DataFrame, dict	Exogenous variable/s included as predictor/s.	None
quantiles	(list, tuple)	Sequence of quantiles to compute, which must be between 0 and 1 inclusive. For example, quantiles of 0.05, 0.5 and 0.95 should be as quantiles = [0.05, 0.5, 0.95].	[0.05, 0.5, 0.95]
n_boot	int	Number of bootstrapping iterations to perform when estimating quantiles.	250
use_in_sample_residuals	bool	If True, residuals from the training data are used as proxy of prediction error to create predictions. If False, out of sample residuals (calibration) are used. Out-of-sample residuals must be precomputed using Forecaster's set_out_sample_residuals() method.	True
use_binned_residuals	bool	If True, residuals are selected based on the predicted values (binned selection). If False, residuals are selected randomly. New in version 0.15.0	True
random_state	int	Seed for the random number generator to ensure reproducibility.	123
suppress_warnings	bool	If True, skforecast warnings will be suppressed during the prediction process. See skforecast.exceptions.warn_skforecast_categories for more information.	False

Returns:

Name	Type	Description
predictions	pandas DataFrame	Long-format DataFrame with the quantiles predicted by the forecaster. For example, if quantiles = [0.05, 0.5, 0.95], the columns are level, q_0.05, q_0.5, q_0.95.

References

.. [1] Forecasting: Principles and Practice (3^rd ed) Rob J Hyndman and George Athanasopoulos. https://otexts.com/fpp3/prediction-intervals.html

Source code in skforecast/recursive/_forecaster_recursive_multiseries.py

@manage_warnings
def predict_quantiles(
    self,
    steps: int,
    levels: str | list[str] | None = None,
    last_window: pd.DataFrame | None = None,
    exog: pd.Series | pd.DataFrame | dict[str, pd.Series | pd.DataFrame] | None = None,
    quantiles: list[float] | tuple[float] = [0.05, 0.5, 0.95],
    n_boot: int = 250,
    use_in_sample_residuals: bool = True,
    use_binned_residuals: bool = True,
    random_state: int = 123,
    suppress_warnings: bool = False
) -> pd.DataFrame:
    """
    Calculate the specified quantiles for each step. After generating 
    multiple forecasting predictions through a bootstrapping process, each 
    quantile is calculated for each step.

    Parameters
    ----------
    steps : int
        Number of steps to predict. 
    levels : str, list, default None
        Time series to be predicted. If `None` all levels whose last window
        ends at the same datetime index will be predicted together.
    last_window : pandas DataFrame, default None
        Series values used to create the predictors (lags) needed in the 
        first iteration of the prediction (t + 1).
        If `last_window = None`, the values stored in `self.last_window_` are
        used to calculate the initial predictors, and the predictions start
        right after training data.
    exog : pandas Series, pandas DataFrame, dict, default None
        Exogenous variable/s included as predictor/s.
    quantiles : list, tuple, default [0.05, 0.5, 0.95]
        Sequence of quantiles to compute, which must be between 0 and 1 
        inclusive. For example, quantiles of 0.05, 0.5 and 0.95 should be as 
        `quantiles = [0.05, 0.5, 0.95]`.
    n_boot : int, default 250
        Number of bootstrapping iterations to perform when estimating quantiles.
    use_in_sample_residuals : bool, default True
        If `True`, residuals from the training data are used as proxy of
        prediction error to create predictions. 
        If `False`, out of sample residuals (calibration) are used. 
        Out-of-sample residuals must be precomputed using Forecaster's
        `set_out_sample_residuals()` method.
    use_binned_residuals : bool, default True
        If `True`, residuals are selected based on the predicted values 
        (binned selection).
        If `False`, residuals are selected randomly.
        **New in version 0.15.0**
    random_state : int, default 123
        Seed for the random number generator to ensure reproducibility.
    suppress_warnings : bool, default False
        If `True`, skforecast warnings will be suppressed during the prediction 
        process. See skforecast.exceptions.warn_skforecast_categories for more
        information.

    Returns
    -------
    predictions : pandas DataFrame
        Long-format DataFrame with the quantiles predicted by the forecaster.
        For example, if `quantiles = [0.05, 0.5, 0.95]`, the columns are
        `level`, `q_0.05`, `q_0.5`, `q_0.95`.

    References
    ----------
    .. [1] Forecasting: Principles and Practice (3rd ed) Rob J Hyndman and George Athanasopoulos.
           https://otexts.com/fpp3/prediction-intervals.html

    """

    check_interval(quantiles=quantiles)

    predictions = self.predict_bootstrapping(
                      steps                   = steps,
                      levels                  = levels,
                      last_window             = last_window,
                      exog                    = exog,
                      n_boot                  = n_boot,
                      use_in_sample_residuals = use_in_sample_residuals,
                      use_binned_residuals    = use_binned_residuals,
                      random_state            = random_state,
                      suppress_warnings       = suppress_warnings
                  )

    quantiles_cols = [f'q_{q}' for q in quantiles]
    predictions[quantiles_cols] = (
        predictions.iloc[:, 1:].quantile(q=quantiles, axis=1).transpose()
    )
    predictions = predictions[['level'] + quantiles_cols]

    return predictions

predict_dist

predict_dist(
    steps,
    distribution,
    levels=None,
    last_window=None,
    exog=None,
    n_boot=250,
    use_in_sample_residuals=True,
    use_binned_residuals=True,
    random_state=123,
    suppress_warnings=False,
)

Fit a given probability distribution for each step. After generating multiple forecasting predictions through a bootstrapping process, each step is fitted to the given distribution.

Parameters:

Name	Type	Description	Default
steps	int	Number of steps to predict.	required
distribution	object	A distribution object from scipy.stats with methods _pdf and fit. For example scipy.stats.norm.	required
levels	(str, list)	Time series to be predicted. If None all levels whose last window ends at the same datetime index will be predicted together.	None
last_window	pandas DataFrame	Series values used to create the predictors (lags) needed in the first iteration of the prediction (t + 1). If last_window = None, the values stored in self.last_window_ are used to calculate the initial predictors, and the predictions start right after training data.	None
exog	pandas Series, pandas DataFrame, dict	Exogenous variable/s included as predictor/s.	None
n_boot	int	Number of bootstrapping iterations to perform when estimating prediction intervals.	250
use_in_sample_residuals	bool	If True, residuals from the training data are used as proxy of prediction error to create predictions. If False, out of sample residuals (calibration) are used. Out-of-sample residuals must be precomputed using Forecaster's set_out_sample_residuals() method.	True
use_binned_residuals	bool	If True, residuals are selected based on the predicted values (binned selection). If False, residuals are selected randomly. New in version 0.15.0	True
random_state	int	Seed for the random number generator to ensure reproducibility.	123
suppress_warnings	bool	If True, skforecast warnings will be suppressed during the prediction process. See skforecast.exceptions.warn_skforecast_categories for more information.	False

Returns:

Name	Type	Description
predictions	pandas DataFrame	Long-format DataFrame with the parameters of the fitted distribution for each step. The columns are level, param_0, param_1, ..., param_n, where param_i are the parameters of the distribution.

References

.. [1] Forecasting: Principles and Practice (3^rd ed) Rob J Hyndman and George Athanasopoulos. https://otexts.com/fpp3/prediction-intervals.html

Source code in skforecast/recursive/_forecaster_recursive_multiseries.py

@manage_warnings
def predict_dist(
    self,
    steps: int,
    distribution: object,
    levels: str | list[str] | None = None,
    last_window: pd.DataFrame | None = None,
    exog: pd.Series | pd.DataFrame | dict[str, pd.Series | pd.DataFrame] | None = None,
    n_boot: int = 250,
    use_in_sample_residuals: bool = True,
    use_binned_residuals: bool = True,
    random_state: int = 123,
    suppress_warnings: bool = False
) -> pd.DataFrame:
    """
    Fit a given probability distribution for each step. After generating 
    multiple forecasting predictions through a bootstrapping process, each 
    step is fitted to the given distribution.

    Parameters
    ----------
    steps : int
        Number of steps to predict. 
    distribution : object
        A distribution object from scipy.stats with methods `_pdf` and `fit`. 
        For example scipy.stats.norm.
    levels : str, list, default None
        Time series to be predicted. If `None` all levels whose last window
        ends at the same datetime index will be predicted together.
    last_window : pandas DataFrame, default None
        Series values used to create the predictors (lags) needed in the 
        first iteration of the prediction (t + 1).
        If `last_window = None`, the values stored in `self.last_window_` are
        used to calculate the initial predictors, and the predictions start
        right after training data.
    exog : pandas Series, pandas DataFrame, dict, default None
        Exogenous variable/s included as predictor/s.
    n_boot : int, default 250
        Number of bootstrapping iterations to perform when estimating prediction
        intervals.
    use_in_sample_residuals : bool, default True
        If `True`, residuals from the training data are used as proxy of
        prediction error to create predictions. 
        If `False`, out of sample residuals (calibration) are used. 
        Out-of-sample residuals must be precomputed using Forecaster's
        `set_out_sample_residuals()` method.
    use_binned_residuals : bool, default True
        If `True`, residuals are selected based on the predicted values 
        (binned selection).
        If `False`, residuals are selected randomly.
        **New in version 0.15.0**
    random_state : int, default 123
        Seed for the random number generator to ensure reproducibility.
    suppress_warnings : bool, default False
        If `True`, skforecast warnings will be suppressed during the prediction 
        process. See skforecast.exceptions.warn_skforecast_categories for more
        information.

    Returns
    -------
    predictions : pandas DataFrame
        Long-format DataFrame with the parameters of the fitted distribution
        for each step. The columns are `level`, `param_0`, `param_1`, ..., 
        `param_n`, where `param_i` are the parameters of the distribution.

    References
    ----------
    .. [1] Forecasting: Principles and Practice (3rd ed) Rob J Hyndman and George Athanasopoulos.
           https://otexts.com/fpp3/prediction-intervals.html

    """

    if not hasattr(distribution, "_pdf") or not callable(getattr(distribution, "fit", None)):
        raise TypeError(
            "`distribution` must be a valid probability distribution object "
            "from scipy.stats, with methods `_pdf` and `fit`."
        )

    predictions = self.predict_bootstrapping(
                      steps                   = steps,
                      levels                  = levels,
                      last_window             = last_window,
                      exog                    = exog,
                      n_boot                  = n_boot,
                      use_in_sample_residuals = use_in_sample_residuals,
                      use_binned_residuals    = use_binned_residuals,
                      random_state            = random_state,
                      suppress_warnings       = suppress_warnings
                  )

    param_names = [
        p for p in inspect.signature(distribution._pdf).parameters 
        if not p == "x"
    ] + ["loc", "scale"]

    predictions[param_names] = (
        predictions.iloc[:, 1:].apply(
            lambda x: distribution.fit(x), axis=1, result_type='expand'
        )
    )
    predictions = predictions[['level'] + param_names]

    return predictions

set_params

set_params(params)

Set new values to the parameters of the scikit-learn model stored in the forecaster. After calling this method, the forecaster is reset to an unfitted state. The fit method must be called before prediction.

Parameters:

Name	Type	Description	Default
params	dict	Parameters values.	required

Returns:

Type	Description
None

Source code in skforecast/recursive/_forecaster_recursive_multiseries.py

def set_params(
    self, 
    params: dict[str, object]
) -> None:
    """
    Set new values to the parameters of the scikit-learn model stored in the
    forecaster. After calling this method, the forecaster is reset to an 
    unfitted state. The `fit` method must be called before prediction.

    Parameters
    ----------
    params : dict
        Parameters values.

    Returns
    -------
    None

    """

    self.estimator = clone(self.estimator)
    self.estimator.set_params(**params)
    self.is_fitted = False

set_lags

set_lags(lags=None)

Set new value to the attribute lags. Attributes lags_names, max_lag and window_size are also updated.

Parameters:

Name	Type	Description	Default
lags	int, list, numpy ndarray, range	Lags used as predictors. Index starts at 1, so lag 1 is equal to t-1. int: include lags from 1 to lags (included). list, 1d numpy ndarray or range: include only lags present in lags, all elements must be int. None: no lags are included as predictors.	None

Returns:

Type	Description
None

Source code in skforecast/recursive/_forecaster_recursive_multiseries.py

def set_lags(
    self, 
    lags: int | list[int] | np.ndarray[int] | range[int] | None = None
) -> None:
    """
    Set new value to the attribute `lags`. Attributes `lags_names`, 
    `max_lag` and `window_size` are also updated.

    Parameters
    ----------
    lags : int, list, numpy ndarray, range, default None
        Lags used as predictors. Index starts at 1, so lag 1 is equal to t-1. 

        - `int`: include lags from 1 to `lags` (included).
        - `list`, `1d numpy ndarray` or `range`: include only lags present in 
        `lags`, all elements must be int.
        - `None`: no lags are included as predictors. 

    Returns
    -------
    None

    """

    if self.window_features is None and lags is None:
        raise ValueError(
            "At least one of the arguments `lags` or `window_features` "
            "must be different from None. This is required to create the "
            "predictors used in training the forecaster."
        )

    self.lags, self.lags_names, self.max_lag = initialize_lags(type(self).__name__, lags)
    self.lags_are_contiguous = (
        self.lags is not None
        and np.array_equal(self.lags, np.arange(1, self.max_lag + 1))
    )
    self.window_size = max(
        [ws for ws in [self.max_lag, self.max_size_window_features] 
         if ws is not None]
    )

    if self.differentiation is not None:
        self.window_size += self.differentiation_max
        if isinstance(self.differentiator, dict):
            for series in self.differentiator.keys():
                if self.differentiator[series] is not None:
                    self.differentiator[series].set_params(window_size=self.window_size)
        else:
            self.differentiator.set_params(window_size=self.window_size)

set_window_features

set_window_features(window_features=None)

Set new value to the attribute window_features. Attributes max_size_window_features, window_features_names, window_features_class_names and window_size are also updated.

Parameters:

Name	Type	Description	Default
window_features	(object, list)	Instance or list of instances used to create window features. Window features are created from the original time series and are included as predictors.	None

Returns:

Type	Description
None

Source code in skforecast/recursive/_forecaster_recursive_multiseries.py

def set_window_features(
    self, 
    window_features: object | list[object] | None = None
) -> None:
    """
    Set new value to the attribute `window_features`. Attributes 
    `max_size_window_features`, `window_features_names`, 
    `window_features_class_names` and `window_size` are also updated.

    Parameters
    ----------
    window_features : object, list, default None
        Instance or list of instances used to create window features. Window features
        are created from the original time series and are included as predictors.

    Returns
    -------
    None

    """

    if window_features is None and self.lags is None:
        raise ValueError(
            "At least one of the arguments `lags` or `window_features` "
            "must be different from None. This is required to create the "
            "predictors used in training the forecaster."
        )

    self.window_features, self.window_features_names, self.max_size_window_features = (
        initialize_window_features(window_features)
    )
    self.window_features_class_names = None
    if window_features is not None:
        self.window_features_class_names = [
            type(wf).__name__ for wf in self.window_features
        ] 
    self.window_size = max(
        [ws for ws in [self.max_lag, self.max_size_window_features] 
         if ws is not None]
    )

    if self.differentiation is not None:
        self.window_size += self.differentiation_max
        if isinstance(self.differentiator, dict):
            for series in self.differentiator.keys():
                if self.differentiator[series] is not None:
                    self.differentiator[series].set_params(window_size=self.window_size)
        else:
            self.differentiator.set_params(window_size=self.window_size)

set_fit_kwargs

set_fit_kwargs(fit_kwargs)

Set new values for the additional keyword arguments passed to the fit method of the estimator.

Parameters:

Name	Type	Description	Default
fit_kwargs	dict	Dict of the form {"argument": new_value}.	required

Returns:

Type	Description
None

Source code in skforecast/recursive/_forecaster_recursive_multiseries.py

def set_fit_kwargs(
    self, 
    fit_kwargs: dict[str, object]
) -> None:
    """
    Set new values for the additional keyword arguments passed to the `fit` 
    method of the estimator.

    Parameters
    ----------
    fit_kwargs : dict
        Dict of the form {"argument": new_value}.

    Returns
    -------
    None

    """

    self.fit_kwargs = check_select_fit_kwargs(self.estimator, fit_kwargs=fit_kwargs)

set_in_sample_residuals

set_in_sample_residuals(
    series,
    exog=None,
    random_state=123,
    suppress_warnings=False,
)

Set in-sample residuals in case they were not calculated during the training process.

In-sample residuals are calculated as the difference between the true values and the predictions made by the forecaster using the training data. The following internal attributes are updated:

• in_sample_residuals_: Dictionary containing a numpy ndarray with the residuals for each series in the form {series: residuals}.
• binner_intervals_: intervals used to bin the residuals are calculated using the quantiles of the predicted values.
• in_sample_residuals_by_bin_: residuals are binned according to the predicted value they are associated with and stored in a dictionary per series, where the keys are the intervals of the predicted values and the values are the residuals associated with that range.

A total of 10_000 residuals are stored in the attribute in_sample_residuals_. If the number of residuals is greater than 10_000, a random sample of 10_000 residuals is stored. The number of residuals stored per bin is limited to 10_000 // self.binner.n_bins_.

Parameters:

Name	Type	Description	Default
series	pandas DataFrame, dict	Training time series.	required
exog	pandas Series, pandas DataFrame, dict	Exogenous variable/s included as predictor/s.	None
random_state	int	Sets a seed to the random sampling for reproducible output.	123
suppress_warnings	bool	If True, skforecast warnings will be suppressed during the sampling process. See skforecast.exceptions.warn_skforecast_categories for more information.	False

Returns:

Type	Description
None

Source code in skforecast/recursive/_forecaster_recursive_multiseries.py

@manage_warnings
def set_in_sample_residuals(
    self,
    series: pd.DataFrame | dict[str, pd.Series | pd.DataFrame],
    exog: pd.Series | pd.DataFrame | dict[str, pd.Series | pd.DataFrame] | None = None,
    random_state: int = 123,
    suppress_warnings: bool = False
) -> None:
    """
    Set in-sample residuals in case they were not calculated during the
    training process. 

    In-sample residuals are calculated as the difference between the true 
    values and the predictions made by the forecaster using the training 
    data. The following internal attributes are updated:

    + `in_sample_residuals_`: Dictionary containing a numpy ndarray with the
    residuals for each series in the form `{series: residuals}`.
    + `binner_intervals_`: intervals used to bin the residuals are calculated
    using the quantiles of the predicted values.
    + `in_sample_residuals_by_bin_`: residuals are binned according to the
    predicted value they are associated with and stored in a dictionary per
    series, where the keys are the intervals of the predicted values and the 
    values are the residuals associated with that range. 

    A total of 10_000 residuals are stored in the attribute `in_sample_residuals_`.
    If the number of residuals is greater than 10_000, a random sample of
    10_000 residuals is stored. The number of residuals stored per bin is
    limited to `10_000 // self.binner.n_bins_`.

    Parameters
    ----------
    series : pandas DataFrame, dict
        Training time series.
    exog : pandas Series, pandas DataFrame, dict, default None
        Exogenous variable/s included as predictor/s.
    random_state : int, default 123
        Sets a seed to the random sampling for reproducible output.
    suppress_warnings : bool, default False
        If `True`, skforecast warnings will be suppressed during the sampling 
        process. See skforecast.exceptions.warn_skforecast_categories for more
        information.

    Returns
    -------
    None

    """

    if not self.is_fitted:
        raise NotFittedError(
            "This forecaster is not fitted yet. Call `fit` with appropriate "
            "arguments before using `set_in_sample_residuals()`."
        )

    (
        X_train,
        y_train,
        series_indexes,
        _,
        X_train_series_names_in_,
        *_
    ) = self._create_train_X_y(
            series=series, exog=exog, store_last_window=False
        )

    # NOTE: Same series names as training is checked in _create_train_X_y.
    series_index_range = {k: v[[0, -1]] for k, v in series_indexes.items()}
    for level in self.training_range_.keys():
        if not series_index_range[level].equals(self.training_range_[level]):
            raise IndexError(
                f"The index range for series '{level}' does not match the range "
                f"used during training. Please ensure the index is aligned "
                f"with the training data.\n"
                f"    Expected : {self.training_range_[level]}\n"
                f"    Received : {series_index_range[level]}"
            )

    X_train_estimator = (
        X_train
        if self.encoding is not None
        else X_train.drop(columns="_level_skforecast")
    )
    X_train_features_names_out_ = X_train_estimator.columns.to_list()
    if not X_train_features_names_out_ == self.X_train_features_names_out_:
        raise ValueError(
            f"Feature mismatch detected after matrix creation. The features "
            f"generated from the provided data do not match those used during "
            f"the training process. To correctly set in-sample residuals, "
            f"ensure that the same data and preprocessing steps are applied.\n"
            f"    Expected output : {self.X_train_features_names_out_}\n"
            f"    Current output  : {X_train_features_names_out_}"
        )

    self.in_sample_residuals_ = {}
    self.in_sample_residuals_by_bin_ = {}
    y_pred = self.estimator.predict(X_train_estimator)
    if self.encoding is not None:
        for level in X_train_series_names_in_:
            if self.encoding == 'onehot':
                mask = X_train[level].to_numpy() == 1.
            else:
                encoded_value = self.encoding_mapping_[level]
                mask = X_train['_level_skforecast'].to_numpy() == encoded_value

            self._binning_in_sample_residuals(
                level                     = level,
                y_true                    = y_train[mask],
                y_pred                    = y_pred[mask],
                store_in_sample_residuals = True,
                random_state              = random_state
            )

    # NOTE: the _unknown_level is a random sample of 10_000 residuals of all levels.
    self._binning_in_sample_residuals(
        level                     = '_unknown_level',
        y_true                    = y_train,
        y_pred                    = y_pred,
        store_in_sample_residuals = True,
        random_state              = random_state
    )

set_out_sample_residuals

set_out_sample_residuals(
    y_true, y_pred, append=False, random_state=123
)

Set new values to the attribute out_sample_residuals_. Out of sample residuals are meant to be calculated using observations that did not participate in the training process. y_true and y_pred are expected to be in the original scale of the time series. Residuals are calculated as y_true - y_pred, after applying the necessary transformations and differentiations if the forecaster includes them (self.transformer_series and self.differentiation).

A total of 10_000 residuals are stored in the attribute out_sample_residuals_. If the number of residuals is greater than 10_000, a random sample of 10_000 residuals is stored.

Parameters:

Name	Type	Description	Default
y_true	dict	Dictionary of numpy ndarrays or pandas Series with the true values of the time series for each series in the form {series: y_true}.	required
y_pred	dict	Dictionary of numpy ndarrays or pandas Series with the predicted values of the time series for each series in the form {series: y_pred}.	required
append	bool	If True, new residuals are added to the once already stored in the attribute out_sample_residuals_. If after appending the new residuals, the limit of 10_000 samples is exceeded, a random sample of 10_000 is kept.	False
random_state	int	Sets a seed to the random sampling for reproducible output.	123

Returns:

Type	Description
None

Notes

Out-of-sample residuals can only be stored for series seen during fit. To save residuals for unseen levels use the key '_unknown_level'. If '_unknown_level' is not provided, or if encoding is None, the residuals of all the levels provided are combined and stored under the key '_unknown_level', binned with the binner fitted for that key.

Source code in skforecast/recursive/_forecaster_recursive_multiseries.py

def set_out_sample_residuals(
    self, 
    y_true: dict[str, np.ndarray | pd.Series],
    y_pred: dict[str, np.ndarray | pd.Series],
    append: bool = False,
    random_state: int = 123
) -> None:
    """
    Set new values to the attribute `out_sample_residuals_`. Out of sample
    residuals are meant to be calculated using observations that did not
    participate in the training process. `y_true` and `y_pred` are expected
    to be in the original scale of the time series. Residuals are calculated
    as `y_true` - `y_pred`, after applying the necessary transformations and
    differentiations if the forecaster includes them (`self.transformer_series`
    and `self.differentiation`).

    A total of 10_000 residuals are stored in the attribute `out_sample_residuals_`.
    If the number of residuals is greater than 10_000, a random sample of
    10_000 residuals is stored.

    Parameters
    ----------
    y_true : dict
        Dictionary of numpy ndarrays or pandas Series with the true values of
        the time series for each series in the form {series: y_true}.
    y_pred : dict
        Dictionary of numpy ndarrays or pandas Series with the predicted values
        of the time series for each series in the form {series: y_pred}.
    append : bool, default False
        If `True`, new residuals are added to the once already stored in the
        attribute `out_sample_residuals_`. If after appending the new residuals,
        the limit of 10_000 samples is exceeded, a random sample of 10_000 is
        kept.
    random_state : int, default 123
        Sets a seed to the random sampling for reproducible output.

    Returns
    -------
    None

    Notes
    -----
    Out-of-sample residuals can only be stored for series seen during 
    fit. To save residuals for unseen levels use the key '_unknown_level'. 
    If '_unknown_level' is not provided, or if `encoding` is `None`, the
    residuals of all the levels provided are combined and stored under the
    key '_unknown_level', binned with the binner fitted for that key.

    """

    if not self.is_fitted:
        raise NotFittedError(
            "This forecaster is not fitted yet. Call `fit` with appropriate "
            "arguments before using `set_out_sample_residuals()`."
        )

    if not isinstance(y_true, dict):
        raise TypeError(
            f"`y_true` must be a dictionary of numpy ndarrays or pandas Series. "
            f"Got {type(y_true)}."
        )

    if not isinstance(y_pred, dict):
        raise TypeError(
            f"`y_pred` must be a dictionary of numpy ndarrays or pandas Series. "
            f"Got {type(y_pred)}."
        )

    if not set(y_true.keys()) == set(y_pred.keys()):
        raise ValueError(
            f"`y_true` and `y_pred` must have the same keys. "
            f"Got {set(y_true.keys())} and {set(y_pred.keys())}."
        )

    for k in y_true.keys():
        if not isinstance(y_true[k], (np.ndarray, pd.Series)):
            raise TypeError(
                f"Values of `y_true` must be numpy ndarrays or pandas Series. "
                f"Got {type(y_true[k])} for series '{k}'."
            )
        if not isinstance(y_pred[k], (np.ndarray, pd.Series)):
            raise TypeError(
                f"Values of `y_pred` must be numpy ndarrays or pandas Series. "
                f"Got {type(y_pred[k])} for series '{k}'."
            )
        if len(y_true[k]) != len(y_pred[k]):
            raise ValueError(
                f"`y_true` and `y_pred` must have the same length. "
                f"Got {len(y_true[k])} and {len(y_pred[k])} for series '{k}'."
            )
        if isinstance(y_true[k], pd.Series) and isinstance(y_pred[k], pd.Series):
            if not y_true[k].index.equals(y_pred[k].index):
                raise ValueError(
                    f"When containing pandas Series, elements in `y_true` and "
                    f"`y_pred` must have the same index. Error with series '{k}'."
                )

    # NOTE: Out-of-sample residuals can only be stored for series seen during 
    # fit. To save residuals for unseen levels use the key '_unknown_level'. 
    series_names_in_ = self.series_names_in_ + ['_unknown_level']
    series_to_update = [
        series for series in series_names_in_ if series in y_pred.keys()
    ]
    if not series_to_update:
        raise ValueError(
            "Provided keys in `y_pred` and `y_true` do not match any series "
            "seen during `fit`. Residuals cannot be updated."
        )

    if self.out_sample_residuals_ is None:
        if self.encoding is not None:
            self.out_sample_residuals_ = {level: None for level in series_names_in_}
            self.out_sample_residuals_by_bin_ = {level: {} for level in series_names_in_}
        else:
            self.out_sample_residuals_ = {'_unknown_level': None}
            self.out_sample_residuals_by_bin_ = {'_unknown_level': {}}

    residuals_by_level = {
        level: self._transform_out_sample_residuals(
                   level  = level,
                   y_true = y_true[level],
                   y_pred = y_pred[level]
               )
        for level in series_to_update
    }

    # NOTE: When '_unknown_level' is not provided, or when encoding is None
    # (no distinction between levels is made), its residuals are the residuals
    # of all the levels provided, binned with the binner of '_unknown_level'.
    # Each level has its own binner, so the bins of a level cannot be reused
    # as the bins of '_unknown_level'.
    if self.encoding is None or '_unknown_level' not in series_to_update:
        if self.encoding is None and list(y_true.keys()) != ['_unknown_level']:
            warnings.warn(
                "As `encoding` is set to `None`, no distinction between levels "
                "is made. All residuals are stored in the '_unknown_level' key.",
                UnknownLevelWarning
            )

        y_pred_all_levels = np.concatenate(
            [y_pred_level for y_pred_level, _ in residuals_by_level.values()]
        )
        residuals_all_levels = np.concatenate(
            [residuals_level for _, residuals_level in residuals_by_level.values()]
        )
        if self.encoding is None:
            residuals_by_level = {}
        residuals_by_level['_unknown_level'] = (
            y_pred_all_levels, residuals_all_levels
        )

    for level, (y_pred_level, residuals_level) in residuals_by_level.items():
        out_sample_residuals, out_sample_residuals_by_bin = (
            self._binning_out_sample_residuals(
                level        = level,
                y_pred       = y_pred_level,
                residuals    = residuals_level,
                append       = append,
                random_state = random_state
            )
        )
        self.out_sample_residuals_[level] = out_sample_residuals
        self.out_sample_residuals_by_bin_[level] = out_sample_residuals_by_bin

get_feature_importances

get_feature_importances(sort_importance=True)

Return feature importances of the estimator stored in the forecaster. Only valid when estimator stores internally the feature importances in the attribute feature_importances_ or coef_.

Parameters:

Name	Type	Description	Default
sort_importance	bool	If True, sorts the feature importances in descending order.	True

Returns:

Name	Type	Description
feature_importances	pandas DataFrame	Feature importances associated with each predictor.

Source code in skforecast/recursive/_forecaster_recursive_multiseries.py

def get_feature_importances(
    self,
    sort_importance: bool = True
) -> pd.DataFrame:
    """
    Return feature importances of the estimator stored in the
    forecaster. Only valid when estimator stores internally the feature
    importances in the attribute `feature_importances_` or `coef_`.

    Parameters
    ----------
    sort_importance: bool, default True
        If `True`, sorts the feature importances in descending order.

    Returns
    -------
    feature_importances : pandas DataFrame
        Feature importances associated with each predictor.

    """

    if not self.is_fitted:
        raise NotFittedError(
            "This forecaster is not fitted yet. Call `fit` with appropriate "
            "arguments before using `get_feature_importances()`."
        )

    if isinstance(self.estimator, Pipeline):
        estimator = self.estimator[-1]
    else:
        estimator = self.estimator

    if hasattr(estimator, 'feature_importances_'):
        feature_importances = estimator.feature_importances_
    elif hasattr(estimator, 'coef_'):
        feature_importances = estimator.coef_
    else:
        warnings.warn(
            f"Impossible to access feature importances for estimator of type "
            f"{type(estimator)}. This method is only valid when the "
            f"estimator stores internally the feature importances in the "
            f"attribute `feature_importances_` or `coef_`."
        )
        feature_importances = None

    if feature_importances is not None:
        feature_importances = pd.DataFrame({
                                  'feature': self.X_train_features_names_out_,
                                  'importance': feature_importances
                              })
        if sort_importance:
            feature_importances = feature_importances.sort_values(
                by='importance', ascending=False
            )

    return feature_importances