ForecasterAutoregDirect¶
ForecasterAutoregDirect (ForecasterBase)
¶
This class turns any regressor compatible with the scikit-learn API into a
autoregressive direct multi-step forecaster. A separate model is created for each forecast time step. See documentation for more details.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
regressor |
object |
An instance of a regressor or pipeline compatible with the scikit-learn API. |
required |
lags |
Union[int, numpy.ndarray, list] |
Lags used as predictors. Index starts at 1, so lag 1 is equal to t-1.
|
required |
steps |
int |
Maximum number of future steps the forecaster will predict when using
method |
required |
transformer_y |
Optional[object] |
An instance of a transformer (preprocessor) compatible with the scikit-learn
preprocessing API with methods: fit, transform, fit_transform and inverse_transform.
ColumnTransformers are not allowed since they do not have inverse_transform method.
The transformation is applied to |
None |
transformer_exog |
Optional[object] |
An instance of a transformer (preprocessor) compatible with the scikit-learn
preprocessing API. The transformation is applied to |
None |
weight_func |
Optional[<built-in function callable>] |
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 |
None |
Attributes:
| Name | Type | Description |
|---|---|---|
regressor |
regressor or pipeline compatible with the scikit-learn API |
An instance of a regressor or pipeline compatible with the scikit-learn API.
One instance of this regressor is trained for each step. All
them are stored in |
regressors_ |
dict |
Dictionary with regressors trained for each step. They are initialized as a copy
of |
steps |
int |
Number of future steps the forecaster will predict when using method
|
lags |
numpy ndarray |
Lags used as predictors. |
transformer_y |
object transformer (preprocessor), default `None` |
An instance of a transformer (preprocessor) compatible with the scikit-learn
preprocessing API with methods: fit, transform, fit_transform and inverse_transform.
ColumnTransformers are not allowed since they do not have inverse_transform method.
The transformation is applied to |
transformer_exog |
object transformer (preprocessor), default `None` |
An instance of a transformer (preprocessor) compatible with the scikit-learn
preprocessing API. The transformation is applied to |
weight_func |
callable |
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 |
source_code_weight_func |
str |
Source code of the custom function used to create weights. New in version 0.6.0 |
max_lag |
int |
Maximum value of lag included in |
window_size |
int |
Size of the window needed to create the predictors. It is equal to
|
last_window |
pandas Series |
Last window the forecaster has seen during trained. It stores the
values needed to predict the next |
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 |
pandas Index |
First and last values of index of the data used during training. |
included_exog |
bool |
If the forecaster has been trained using exogenous variable/s. |
exog_type |
type |
Type of exogenous variable/s used in training. |
exog_col_names |
list |
Names of columns of |
X_train_col_names |
list |
Names of columns of the matrix created internally for training. |
fitted |
Bool |
Tag to identify if the regressor has been fitted (trained). |
creation_date |
str |
Date of creation. |
fit_date |
str |
Date of last fit. |
skforcast_version |
str |
Version of skforecast library used to create the forecaster. |
python_version |
str |
Version of python used to create the forecaster. New in version 0.5.0 |
Source code in skforecast/ForecasterAutoregDirect/ForecasterAutoregDirect.py
class ForecasterAutoregDirect(ForecasterBase):
"""
This class turns any regressor compatible with the scikit-learn API into a
autoregressive direct multi-step forecaster. A separate model is created for
each forecast time step. See documentation for more details.
Parameters
----------
regressor : regressor or pipeline compatible with the scikit-learn API
An instance of a regressor or pipeline compatible with the scikit-learn API.
lags : int, list, 1d 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`, `numpy ndarray` or range: include only lags present in `lags`.
steps : int
Maximum number of future steps the forecaster will predict when using
method `predict()`. Since a different model is created for each step,
this value should be defined before training.
transformer_y : object transformer (preprocessor), default `None`
An instance of a transformer (preprocessor) compatible with the scikit-learn
preprocessing API with methods: fit, transform, fit_transform and inverse_transform.
ColumnTransformers are not allowed since they do not have inverse_transform method.
The transformation is applied to `y` before training the forecaster.
**New in version 0.5.0**
transformer_exog : object transformer (preprocessor), default `None`
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.
**New in version 0.5.0**
weight_func : callable, default `None`
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 `regressor` does not have the argument `sample_weight` in its `fit`
method. The resulting `sample_weight` cannot have negative values.
**New in version 0.6.0**
Attributes
----------
regressor : regressor or pipeline compatible with the scikit-learn API
An instance of a regressor or pipeline compatible with the scikit-learn API.
One instance of this regressor is trained for each step. All
them are stored in `self.regressors_`.
regressors_ : dict
Dictionary with regressors trained for each step. They are initialized as a copy
of `regressor`.
steps : int
Number of future steps the forecaster will predict when using method
`predict()`. Since a different model is created for each step, this value
should be defined before training.
lags : numpy ndarray
Lags used as predictors.
transformer_y : object transformer (preprocessor), default `None`
An instance of a transformer (preprocessor) compatible with the scikit-learn
preprocessing API with methods: fit, transform, fit_transform and inverse_transform.
ColumnTransformers are not allowed since they do not have inverse_transform method.
The transformation is applied to `y` before training the forecaster.
**New in version 0.5.0**
transformer_exog : object transformer (preprocessor), default `None`
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.
**New in version 0.5.0**
weight_func : callable
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 `regressor` does not have the argument `sample_weight` in its `fit`
method.
**New in version 0.6.0**
source_code_weight_func : str
Source code of the custom function used to create weights.
**New in version 0.6.0**
max_lag : int
Maximum value of lag included in `lags`.
window_size : int
Size of the window needed to create the predictors. It is equal to
`max_lag`.
last_window : pandas Series
Last window the forecaster has seen during trained. It stores the
values needed to predict the next `step` right after the training data.
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 : pandas Index
First and last values of index of the data used during training.
included_exog : bool
If the forecaster has been trained using exogenous variable/s.
exog_type : type
Type of exogenous variable/s used in training.
exog_col_names : list
Names of columns of `exog` if `exog` used in training was a pandas
DataFrame.
X_train_col_names : list
Names of columns of the matrix created internally for training.
fitted : Bool
Tag to identify if the regressor has been fitted (trained).
creation_date: str
Date of creation.
fit_date : str
Date of last fit.
skforcast_version : str
Version of skforecast library used to create the forecaster.
python_version : str
Version of python used to create the forecaster.
**New in version 0.5.0**
Notes
-----
A separate model is created for each forecast time step. It is important to
note that all models share the same configuration of parameters and
hyperparameters.
"""
def __init__(
self,
regressor: object,
steps: int,
lags: Union[int, np.ndarray, list],
transformer_y: Optional[object]=None,
transformer_exog: Optional[object]=None,
weight_func: Optional[callable]=None
) -> None:
self.regressor = regressor
self.steps = steps
self.transformer_y = transformer_y
self.transformer_exog = transformer_exog
self.weight_func = weight_func
self.source_code_weight_func = None
self.index_type = None
self.index_freq = None
self.training_range = None
self.last_window = None
self.included_exog = False
self.exog_type = None
self.exog_col_names = None
self.X_train_col_names = None
self.fitted = False
self.creation_date = pd.Timestamp.today().strftime('%Y-%m-%d %H:%M:%S')
self.fit_date = None
self.skforcast_version = skforecast.__version__
self.python_version = sys.version.split(" ")[0]
if not isinstance(steps, int):
raise TypeError(
f"`steps` argument must be an int greater than or equal to 1. "
f"Got {type(steps)}."
)
if steps < 1:
raise ValueError(
f"`steps` argument must be greater than or equal to 1. Got {steps}."
)
self.regressors_ = {step: clone(self.regressor) for step in range(steps)}
self.lags = initialize_lags(type(self).__name__, lags)
self.max_lag = max(self.lags)
self.window_size = self.max_lag
self.weight_func, self.source_code_weight_func, _ = initialize_weights(
forecaster_type = type(self).__name__,
regressor = regressor,
weight_func = weight_func,
series_weights = None
)
def __repr__(
self
) -> str:
"""
Information displayed when a ForecasterAutoregDirect object is printed.
"""
if isinstance(self.regressor, sklearn.pipeline.Pipeline):
name_pipe_steps = tuple(name + "__" for name in self.regressor.named_steps.keys())
params = {key : value for key, value in self.regressor.get_params().items() \
if key.startswith(name_pipe_steps)}
else:
params = self.regressor.get_params()
info = (
f"{'=' * len(str(type(self)).split('.')[1])} \n"
f"{str(type(self)).split('.')[1]} \n"
f"{'=' * len(str(type(self)).split('.')[1])} \n"
f"Regressor: {self.regressor} \n"
f"Lags: {self.lags} \n"
f"Transformer for y: {self.transformer_y} \n"
f"Transformer for exog: {self.transformer_exog} \n"
f"Included weights function: {True if self.weight_func is not None else False} \n"
f"Window size: {self.window_size} \n"
f"Maximum steps predicted: {self.steps} \n"
f"Weight function included: {True if self.weight_func is not None else False} \n"
f"Exogenous included: {self.included_exog} \n"
f"Type of exogenous variable: {self.exog_type} \n"
f"Exogenous variables names: {self.exog_col_names} \n"
f"Training range: {self.training_range.to_list() if self.fitted else None} \n"
f"Training index type: {str(self.index_type).split('.')[-1][:-2] if self.fitted else None} \n"
f"Training index frequency: {self.index_freq if self.fitted else None} \n"
f"Regressor parameters: {params} \n"
f"Creation date: {self.creation_date} \n"
f"Last fit date: {self.fit_date} \n"
f"Skforecast version: {self.skforcast_version} \n"
f"Python version: {self.python_version} \n"
)
return info
def _create_lags(
self,
y: np.ndarray
) -> Tuple[np.ndarray, np.ndarray]:
"""
Transforms a 1d array into a 2d array (X) and a 1d array (y). Each row
in X is associated with a value of y and it represents the lags that
precede it.
Notice that, the returned matrix X_data, contains the lag 1 in the first
column, the lag 2 in the second column and so on.
Parameters
----------
y : 1d numpy ndarray
Training time series.
Returns
-------
X_data : 2d numpy ndarray, shape (samples - max(self.lags), len(self.lags))
2d numpy array with the lagged values (predictors).
y_data : 1d numpy ndarray, shape (samples - max(self.lags),)
Values of the time series related to each row of `X_data`.
"""
n_splits = len(y) - self.max_lag - (self.steps - 1) # rows of y_data
if n_splits <= 0:
raise ValueError(
f'The maximum lag ({self.max_lag}) must be less than the length '
f'of the series minus the number of steps ({len(y)-(self.steps-1)}).'
)
X_data = np.full(shape=(n_splits, len(self.lags)), fill_value=np.nan, dtype=float)
for i, lag in enumerate(self.lags):
X_data[:, i] = y[self.max_lag - lag : -(lag + self.steps - 1)]
y_data = np.full(shape=(n_splits, self.steps), fill_value=np.nan, dtype=float)
for step in range(self.steps):
y_data[:, step] = y[self.max_lag + step : self.max_lag + step + n_splits]
return X_data, y_data
def create_train_X_y(
self,
y: pd.Series,
exog: Optional[Union[pd.Series, pd.DataFrame]]=None
) -> Tuple[pd.DataFrame, pd.DataFrame]:
"""
Create training matrices from univariate time series and exogenous
variables. The resulting matrices contain the target variable and predictors
needed to train all the regressors (one per step).
Parameters
----------
y : pandas Series
Training time series.
exog : pandas Series, pandas DataFrame, default `None`
Exogenous variable/s included as predictor/s. Must have the same
number of observations as `y` and their indexes must be aligned.
Returns
-------
X_train : pandas DataFrame, shape (len(y) - self.max_lag, len(self.lags) + exog.shape[1]*steps)
Pandas DataFrame with the training values (predictors) for each step.
y_train : pandas DataFrame, shape (len(y) - self.max_lag, )
Values (target) of the time series related to each row of `X_train`
for each step.
"""
if len(y) < self.max_lag + self.steps:
raise ValueError(
f'Minimum length of `y` for training this forecaster is '
f'{self.max_lag + self.steps}. Got {len(y)}.'
)
check_y(y=y)
y = transform_series(
series = y,
transformer = self.transformer_y,
fit = True,
inverse_transform = False
)
y_values, y_index = preprocess_y(y=y)
X_train, y_train = self._create_lags(y=y_values)
y_train_col_names = [f"y_step_{i+1}" for i in range(self.steps)]
X_train_col_names = [f"lag_{i}" for i in self.lags]
if exog is not None:
if len(exog) != len(y):
raise ValueError(
f'`exog` must have same number of samples as `y`. '
f'length `exog`: ({len(exog)}), length `y`: ({len(y)})'
)
check_exog(exog=exog)
# Need here for filter_train_X_y_for_step to work without fitting
self.included_exog = True
if isinstance(exog, pd.Series):
exog = transform_series(
series = exog,
transformer = self.transformer_exog,
fit = True,
inverse_transform = False
)
else:
exog = transform_dataframe(
df = exog,
transformer = self.transformer_exog,
fit = True,
inverse_transform = False
)
exog_values, exog_index = preprocess_exog(exog=exog)
if not (exog_index[:len(y_index)] == y_index).all():
raise ValueError(
('Different index for `y` and `exog`. They must be equal '
'to ensure the correct alignment of values.')
)
col_names_exog = exog.columns if isinstance(exog, pd.DataFrame) else [exog.name]
# Transform exog to match direct format
X_exog = exog_to_direct(exog=exog_values, steps=self.steps)
col_names_exog = [f"{col_name}_step_{i+1}" for col_name in col_names_exog for i in range(self.steps)]
X_train_col_names.extend(col_names_exog)
# The first `self.max_lag` positions have to be removed from X_exog
# since they are not in X_lags.
X_exog = X_exog[-X_train.shape[0]:, ]
X_train = np.column_stack((X_train, X_exog))
X_train = pd.DataFrame(
data = X_train,
columns = X_train_col_names,
index = y_index[self.max_lag + (self.steps -1): ]
)
self.X_train_col_names = X_train_col_names
y_train = pd.DataFrame(
data = y_train,
index = y_index[self.max_lag + (self.steps -1): ],
columns = y_train_col_names,
)
return X_train, y_train
def filter_train_X_y_for_step(
self,
step: int,
X_train: pd.DataFrame,
y_train: pd.Series
) -> Tuple[pd.DataFrame, pd.Series]:
"""
Select columns needed to train a forecaster for a specific step. The input
matrices should be created with created with `create_train_X_y()`.
Parameters
----------
step : int
step for which columns must be selected selected. Starts at 1.
X_train : pandas DataFrame
Pandas DataFrame with the training values (predictors).
y_train : pandas Series
Values (target) of the time series related to each row of `X_train`.
Returns
-------
X_train_step : pandas DataFrame
Pandas DataFrame with the training values (predictors) for step.
y_train_step : pandas Series, shape (len(y) - self.max_lag)
Values (target) of the time series related to each row of `X_train`.
"""
if (step < 1) or (step > self.steps):
raise ValueError(
f"Invalid value `step`. For this forecaster, minimum value is 1 "
f"and the maximum step is {self.steps}."
)
step = step - 1 # To start at 0
y_train_step = y_train.iloc[:, step]
if not self.included_exog:
X_train_step = X_train
else:
idx_columns_lags = np.arange(len(self.lags))
idx_columns_exog = np.arange(X_train.shape[1])[len(self.lags) + step::self.steps]
idx_columns = np.hstack((idx_columns_lags, idx_columns_exog))
X_train_step = X_train.iloc[:, idx_columns]
return X_train_step, y_train_step
def create_sample_weights(
self,
X_train: pd.DataFrame,
)-> np.ndarray:
"""
Crate weights for each observation according to the forecaster's attribute
`weight_func`.
Parameters
----------
X_train : pandas DataFrame
Dataframe generated with the methods `create_train_X_y` and
`filter_train_X_y_for_step`, first return.
Returns
-------
sample_weight : numpy ndarray
Weights to use in `fit` method.
"""
sample_weight = None
if self.weight_func is not None:
sample_weight = self.weight_func(X_train.index)
if sample_weight is not None:
if np.isnan(sample_weight).any():
raise ValueError(
"The resulting `sample_weight` cannot have NaN values."
)
if np.any(sample_weight < 0):
raise ValueError(
"The resulting `sample_weight` cannot have negative values."
)
if np.sum(sample_weight) == 0:
raise ValueError(
("The resulting `sample_weight` cannot be normalized because "
"the sum of the weights is zero.")
)
return sample_weight
def fit(
self,
y: pd.Series,
exog: Optional[Union[pd.Series, pd.DataFrame]]=None
) -> None:
"""
Training Forecaster.
Parameters
----------
y : pandas Series
Training time series.
exog : pandas Series, pandas DataFrame, default `None`
Exogenous variable/s included as predictor/s. Must have the same
number of observations as `y` and their indexes must be aligned so
that y[i] is regressed on exog[i].
Returns
-------
None
"""
# Reset values in case the forecaster has already been fitted.
self.index_type = None
self.index_freq = None
self.last_window = None
self.included_exog = False
self.exog_type = None
self.exog_col_names = None
self.X_train_col_names = None
self.fitted = False
self.training_range = None
if exog is not None:
self.exog_type = type(exog)
self.exog_col_names = \
exog.columns.to_list() if isinstance(exog, pd.DataFrame) else exog.name
X_train, y_train = self.create_train_X_y(y=y, exog=exog)
# Train one regressor for each step
for step in range(self.steps):
X_train_step, y_train_step = self.filter_train_X_y_for_step(
step = step + 1,
X_train = X_train,
y_train = y_train
)
sample_weight = self.create_sample_weights(X_train=X_train_step)
if sample_weight is not None:
if not str(type(self.regressor)) == "<class 'xgboost.sklearn.XGBRegressor'>":
self.regressors_[step].fit(
X = X_train_step,
y = y_train_step,
sample_weight = sample_weight
)
else:
self.regressors_[step].fit(
X = X_train_step.to_numpy(),
y = y_train_step.to_numpy(),
sample_weight = sample_weight
)
else:
if not str(type(self.regressor)) == "<class 'xgboost.sklearn.XGBRegressor'>":
self.regressors_[step].fit(X=X_train_step, y=y_train_step)
else:
self.regressors_[step].fit(X=X_train_step.to_numpy(), y=y_train_step.to_numpy())
self.fitted = True
self.fit_date = pd.Timestamp.today().strftime('%Y-%m-%d %H:%M:%S')
self.training_range = preprocess_y(y=y)[1][[0, -1]]
self.index_type = type(X_train.index)
if isinstance(X_train.index, pd.DatetimeIndex):
self.index_freq = X_train.index.freqstr
else:
self.index_freq = X_train.index.step
self.last_window = y.iloc[-self.max_lag:].copy()
def predict(
self,
steps: Optional[Union[int, list]]=None,
last_window: Optional[pd.Series]=None,
exog: Optional[Union[pd.Series, pd.DataFrame]]=None
) -> pd.Series:
"""
Predict n steps ahead.
Parameters
----------
steps : int, list, None, default `None`
Predict n steps. The value of `steps` must be less than or equal to the
value of steps defined when initializing the forecaster. Starts at 1.
If int:
Only steps within the range of 1 to int are predicted.
If list:
List of ints. Only the steps contained in the list are predicted.
If `None`:
As many steps are predicted as were defined at initialization.
last_window : pandas Series, default `None`
Values of the series used to create the predictors (lags) need in the
first iteration of 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.
Returns
-------
predictions : pandas Series
Predicted values.
"""
if isinstance(steps, int):
steps = list(range(steps))
elif steps is None:
steps = list(range(self.steps))
elif isinstance(steps, list):
steps = list(np.array(steps) - 1) # To start at 0 for indexing
for step in steps:
if not isinstance(step, (int, np.int64, np.int32)):
raise TypeError(
f"`steps` argument must be an int, a list of ints or `None`. "
f"Got {type(steps)}."
)
check_predict_input(
forecaster_type = type(self).__name__,
steps = steps,
fitted = self.fitted,
included_exog = self.included_exog,
index_type = self.index_type,
index_freq = self.index_freq,
window_size = self.window_size,
last_window = last_window,
exog = exog,
exog_type = self.exog_type,
exog_col_names = self.exog_col_names,
interval = None,
max_steps = self.steps,
levels = None,
series_col_names = None
)
if exog is not None:
if isinstance(exog, pd.DataFrame):
exog = transform_dataframe(
df = exog,
transformer = self.transformer_exog,
fit = False,
inverse_transform = False
)
else:
exog = transform_series(
series = exog,
transformer = self.transformer_exog,
fit = False,
inverse_transform = False
)
exog_values, _ = preprocess_exog(
exog = exog.iloc[:max(steps)+1, ]
)
exog_values = exog_to_direct(exog=exog_values, steps=max(steps)+1)
else:
exog_values = None
if last_window is None:
last_window = self.last_window.copy()
last_window = transform_series(
series = last_window,
transformer = self.transformer_y,
fit = False,
inverse_transform = False
)
last_window_values, last_window_index = preprocess_last_window(
last_window = last_window
)
X_lags = last_window_values[-self.lags].reshape(1, -1)
predictions = np.full(shape=len(steps), fill_value=np.nan)
for i, step in enumerate(steps):
regressor = self.regressors_[step]
if exog is None:
X = X_lags
else:
# Only columns from exog related with the current step are selected.
X = np.hstack([X_lags, exog_values[0][step::max(steps)+1].reshape(1, -1)])
with warnings.catch_warnings():
# Suppress scikit-learn warning: "X does not have valid feature names,
# but NoOpTransformer was fitted with feature names".
warnings.simplefilter("ignore")
predictions[i] = regressor.predict(X)
idx = expand_index(index=last_window_index, steps=max(steps)+1)
predictions = pd.Series(
data = predictions.reshape(-1),
index = idx[steps],
name = 'pred'
)
predictions = transform_series(
series = predictions,
transformer = self.transformer_y,
fit = False,
inverse_transform = True
)
return predictions
def set_params(
self,
**params: dict
) -> None:
"""
Set new values to the parameters of the scikit learn model stored in the
forecaster. It is important to note that all models share the same
configuration of parameters and hyperparameters.
Parameters
----------
params : dict
Parameters values.
Returns
-------
self
"""
self.regressor = clone(self.regressor)
self.regressor.set_params(**params)
self.regressors_ = {step: clone(self.regressor) for step in range(self.steps)}
def set_lags(
self,
lags: Union[int, list, np.ndarray, range]
) -> None:
"""
Set new value to the attribute `lags`.
Attributes `max_lag` and `window_size` are also updated.
Parameters
----------
lags : int, list, 1D np.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`.
`list` or `np.ndarray`: include only lags present in `lags`.
Returns
-------
None
"""
self.lags = initialize_lags(type(self).__name__, lags)
self.max_lag = max(self.lags)
self.window_size = max(self.lags)
def get_feature_importance(
self,
step: int
) -> pd.DataFrame:
"""
Return impurity-based feature importance of the model stored in
the forecaster for a specific step. Since a separate model is created for
each forecast time step, it is necessary to select the model from which
retrieve information. Only valid when regressor stores internally the
feature importance in the attribute `feature_importances_` or `coef_`.
Parameters
----------
step : int
Model from which retrieve information (a separate model is created
for each forecast time step). First step is 1.
Returns
-------
feature_importance : pandas DataFrame
Feature importance associated with each predictor.
"""
if not isinstance(step, int):
raise TypeError(
f'`step` must be an integer. Got {type(step)}.'
)
if self.fitted == False:
raise sklearn.exceptions.NotFittedError(
"This forecaster is not fitted yet. Call `fit` with appropriate "
"arguments before using `get_feature_importance()`."
)
if (step < 1) or (step > self.steps):
raise ValueError(
f"The step must have a value from 1 to the maximum number of steps "
f"({self.steps}). Got {step}."
)
# Stored regressors start at index 0
step = step - 1
if isinstance(self.regressor, sklearn.pipeline.Pipeline):
estimator = self.regressors_[step][-1]
else:
estimator = self.regressors_[step]
idx_columns_lags = np.arange(len(self.lags))
idx_columns_exog = np.array([], dtype=int)
if self.included_exog:
idx_columns_exog = np.arange(len(self.X_train_col_names))[len(self.lags) + step::self.steps]
idx_columns = np.hstack((idx_columns_lags, idx_columns_exog))
feature_names = [self.X_train_col_names[i] for i in idx_columns]
feature_names = [name.replace(f"_step_{step+1}", "") for name in feature_names]
try:
feature_importance = pd.DataFrame({
'feature': feature_names,
'importance' : estimator.feature_importances_
})
except:
try:
feature_importance = pd.DataFrame({
'feature': feature_names,
'importance' : estimator.coef_
})
except:
warnings.warn(
f"Impossible to access feature importance for regressor of type "
f"{type(estimator)}. This method is only valid when the "
f"regressor stores internally the feature importance in the "
f"attribute `feature_importances_` or `coef_`."
)
feature_importance = None
return feature_importance
create_sample_weights(self, X_train)
¶
Crate weights for each observation according to the forecaster's attribute
weight_func.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
X_train |
DataFrame |
Dataframe generated with the methods |
required |
Returns:
| Type | Description |
|---|---|
ndarray |
Weights to use in |
Source code in skforecast/ForecasterAutoregDirect/ForecasterAutoregDirect.py
def create_sample_weights(
self,
X_train: pd.DataFrame,
)-> np.ndarray:
"""
Crate weights for each observation according to the forecaster's attribute
`weight_func`.
Parameters
----------
X_train : pandas DataFrame
Dataframe generated with the methods `create_train_X_y` and
`filter_train_X_y_for_step`, first return.
Returns
-------
sample_weight : numpy ndarray
Weights to use in `fit` method.
"""
sample_weight = None
if self.weight_func is not None:
sample_weight = self.weight_func(X_train.index)
if sample_weight is not None:
if np.isnan(sample_weight).any():
raise ValueError(
"The resulting `sample_weight` cannot have NaN values."
)
if np.any(sample_weight < 0):
raise ValueError(
"The resulting `sample_weight` cannot have negative values."
)
if np.sum(sample_weight) == 0:
raise ValueError(
("The resulting `sample_weight` cannot be normalized because "
"the sum of the weights is zero.")
)
return sample_weight
create_train_X_y(self, y, exog=None)
¶
Create training matrices from univariate time series and exogenous
variables. The resulting matrices contain the target variable and predictors needed to train all the regressors (one per step).
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
y |
Series |
Training time series. |
required |
exog |
Union[pandas.core.series.Series, pandas.core.frame.DataFrame] |
Exogenous variable/s included as predictor/s. Must have the same
number of observations as |
None |
Returns:
| Type | Description |
|---|---|
Tuple[pandas.core.frame.DataFrame, pandas.core.frame.DataFrame] |
Pandas DataFrame with the training values (predictors) for each step. |
Source code in skforecast/ForecasterAutoregDirect/ForecasterAutoregDirect.py
def create_train_X_y(
self,
y: pd.Series,
exog: Optional[Union[pd.Series, pd.DataFrame]]=None
) -> Tuple[pd.DataFrame, pd.DataFrame]:
"""
Create training matrices from univariate time series and exogenous
variables. The resulting matrices contain the target variable and predictors
needed to train all the regressors (one per step).
Parameters
----------
y : pandas Series
Training time series.
exog : pandas Series, pandas DataFrame, default `None`
Exogenous variable/s included as predictor/s. Must have the same
number of observations as `y` and their indexes must be aligned.
Returns
-------
X_train : pandas DataFrame, shape (len(y) - self.max_lag, len(self.lags) + exog.shape[1]*steps)
Pandas DataFrame with the training values (predictors) for each step.
y_train : pandas DataFrame, shape (len(y) - self.max_lag, )
Values (target) of the time series related to each row of `X_train`
for each step.
"""
if len(y) < self.max_lag + self.steps:
raise ValueError(
f'Minimum length of `y` for training this forecaster is '
f'{self.max_lag + self.steps}. Got {len(y)}.'
)
check_y(y=y)
y = transform_series(
series = y,
transformer = self.transformer_y,
fit = True,
inverse_transform = False
)
y_values, y_index = preprocess_y(y=y)
X_train, y_train = self._create_lags(y=y_values)
y_train_col_names = [f"y_step_{i+1}" for i in range(self.steps)]
X_train_col_names = [f"lag_{i}" for i in self.lags]
if exog is not None:
if len(exog) != len(y):
raise ValueError(
f'`exog` must have same number of samples as `y`. '
f'length `exog`: ({len(exog)}), length `y`: ({len(y)})'
)
check_exog(exog=exog)
# Need here for filter_train_X_y_for_step to work without fitting
self.included_exog = True
if isinstance(exog, pd.Series):
exog = transform_series(
series = exog,
transformer = self.transformer_exog,
fit = True,
inverse_transform = False
)
else:
exog = transform_dataframe(
df = exog,
transformer = self.transformer_exog,
fit = True,
inverse_transform = False
)
exog_values, exog_index = preprocess_exog(exog=exog)
if not (exog_index[:len(y_index)] == y_index).all():
raise ValueError(
('Different index for `y` and `exog`. They must be equal '
'to ensure the correct alignment of values.')
)
col_names_exog = exog.columns if isinstance(exog, pd.DataFrame) else [exog.name]
# Transform exog to match direct format
X_exog = exog_to_direct(exog=exog_values, steps=self.steps)
col_names_exog = [f"{col_name}_step_{i+1}" for col_name in col_names_exog for i in range(self.steps)]
X_train_col_names.extend(col_names_exog)
# The first `self.max_lag` positions have to be removed from X_exog
# since they are not in X_lags.
X_exog = X_exog[-X_train.shape[0]:, ]
X_train = np.column_stack((X_train, X_exog))
X_train = pd.DataFrame(
data = X_train,
columns = X_train_col_names,
index = y_index[self.max_lag + (self.steps -1): ]
)
self.X_train_col_names = X_train_col_names
y_train = pd.DataFrame(
data = y_train,
index = y_index[self.max_lag + (self.steps -1): ],
columns = y_train_col_names,
)
return X_train, y_train
filter_train_X_y_for_step(self, step, X_train, y_train)
¶
Select columns needed to train a forecaster for a specific step. The input
matrices should be created with created with create_train_X_y().
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
step |
int |
step for which columns must be selected selected. Starts at 1. |
required |
X_train |
DataFrame |
Pandas DataFrame with the training values (predictors). |
required |
y_train |
Series |
Values (target) of the time series related to each row of |
required |
Returns:
| Type | Description |
|---|---|
Tuple[pandas.core.frame.DataFrame, pandas.core.series.Series] |
Pandas DataFrame with the training values (predictors) for step. |
Source code in skforecast/ForecasterAutoregDirect/ForecasterAutoregDirect.py
def filter_train_X_y_for_step(
self,
step: int,
X_train: pd.DataFrame,
y_train: pd.Series
) -> Tuple[pd.DataFrame, pd.Series]:
"""
Select columns needed to train a forecaster for a specific step. The input
matrices should be created with created with `create_train_X_y()`.
Parameters
----------
step : int
step for which columns must be selected selected. Starts at 1.
X_train : pandas DataFrame
Pandas DataFrame with the training values (predictors).
y_train : pandas Series
Values (target) of the time series related to each row of `X_train`.
Returns
-------
X_train_step : pandas DataFrame
Pandas DataFrame with the training values (predictors) for step.
y_train_step : pandas Series, shape (len(y) - self.max_lag)
Values (target) of the time series related to each row of `X_train`.
"""
if (step < 1) or (step > self.steps):
raise ValueError(
f"Invalid value `step`. For this forecaster, minimum value is 1 "
f"and the maximum step is {self.steps}."
)
step = step - 1 # To start at 0
y_train_step = y_train.iloc[:, step]
if not self.included_exog:
X_train_step = X_train
else:
idx_columns_lags = np.arange(len(self.lags))
idx_columns_exog = np.arange(X_train.shape[1])[len(self.lags) + step::self.steps]
idx_columns = np.hstack((idx_columns_lags, idx_columns_exog))
X_train_step = X_train.iloc[:, idx_columns]
return X_train_step, y_train_step
fit(self, y, exog=None)
¶
Training Forecaster.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
y |
Series |
Training time series. |
required |
exog |
Union[pandas.core.series.Series, pandas.core.frame.DataFrame] |
Exogenous variable/s included as predictor/s. Must have the same
number of observations as |
None |
Source code in skforecast/ForecasterAutoregDirect/ForecasterAutoregDirect.py
def fit(
self,
y: pd.Series,
exog: Optional[Union[pd.Series, pd.DataFrame]]=None
) -> None:
"""
Training Forecaster.
Parameters
----------
y : pandas Series
Training time series.
exog : pandas Series, pandas DataFrame, default `None`
Exogenous variable/s included as predictor/s. Must have the same
number of observations as `y` and their indexes must be aligned so
that y[i] is regressed on exog[i].
Returns
-------
None
"""
# Reset values in case the forecaster has already been fitted.
self.index_type = None
self.index_freq = None
self.last_window = None
self.included_exog = False
self.exog_type = None
self.exog_col_names = None
self.X_train_col_names = None
self.fitted = False
self.training_range = None
if exog is not None:
self.exog_type = type(exog)
self.exog_col_names = \
exog.columns.to_list() if isinstance(exog, pd.DataFrame) else exog.name
X_train, y_train = self.create_train_X_y(y=y, exog=exog)
# Train one regressor for each step
for step in range(self.steps):
X_train_step, y_train_step = self.filter_train_X_y_for_step(
step = step + 1,
X_train = X_train,
y_train = y_train
)
sample_weight = self.create_sample_weights(X_train=X_train_step)
if sample_weight is not None:
if not str(type(self.regressor)) == "<class 'xgboost.sklearn.XGBRegressor'>":
self.regressors_[step].fit(
X = X_train_step,
y = y_train_step,
sample_weight = sample_weight
)
else:
self.regressors_[step].fit(
X = X_train_step.to_numpy(),
y = y_train_step.to_numpy(),
sample_weight = sample_weight
)
else:
if not str(type(self.regressor)) == "<class 'xgboost.sklearn.XGBRegressor'>":
self.regressors_[step].fit(X=X_train_step, y=y_train_step)
else:
self.regressors_[step].fit(X=X_train_step.to_numpy(), y=y_train_step.to_numpy())
self.fitted = True
self.fit_date = pd.Timestamp.today().strftime('%Y-%m-%d %H:%M:%S')
self.training_range = preprocess_y(y=y)[1][[0, -1]]
self.index_type = type(X_train.index)
if isinstance(X_train.index, pd.DatetimeIndex):
self.index_freq = X_train.index.freqstr
else:
self.index_freq = X_train.index.step
self.last_window = y.iloc[-self.max_lag:].copy()
get_feature_importance(self, step)
¶
Return impurity-based feature importance of the model stored in
the forecaster for a specific step. Since a separate model is created for
each forecast time step, it is necessary to select the model from which
retrieve information. Only valid when regressor stores internally the
feature importance in the attribute feature_importances_ or coef_.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
step |
int |
Model from which retrieve information (a separate model is created for each forecast time step). First step is 1. |
required |
Returns:
| Type | Description |
|---|---|
DataFrame |
Feature importance associated with each predictor. |
Source code in skforecast/ForecasterAutoregDirect/ForecasterAutoregDirect.py
def get_feature_importance(
self,
step: int
) -> pd.DataFrame:
"""
Return impurity-based feature importance of the model stored in
the forecaster for a specific step. Since a separate model is created for
each forecast time step, it is necessary to select the model from which
retrieve information. Only valid when regressor stores internally the
feature importance in the attribute `feature_importances_` or `coef_`.
Parameters
----------
step : int
Model from which retrieve information (a separate model is created
for each forecast time step). First step is 1.
Returns
-------
feature_importance : pandas DataFrame
Feature importance associated with each predictor.
"""
if not isinstance(step, int):
raise TypeError(
f'`step` must be an integer. Got {type(step)}.'
)
if self.fitted == False:
raise sklearn.exceptions.NotFittedError(
"This forecaster is not fitted yet. Call `fit` with appropriate "
"arguments before using `get_feature_importance()`."
)
if (step < 1) or (step > self.steps):
raise ValueError(
f"The step must have a value from 1 to the maximum number of steps "
f"({self.steps}). Got {step}."
)
# Stored regressors start at index 0
step = step - 1
if isinstance(self.regressor, sklearn.pipeline.Pipeline):
estimator = self.regressors_[step][-1]
else:
estimator = self.regressors_[step]
idx_columns_lags = np.arange(len(self.lags))
idx_columns_exog = np.array([], dtype=int)
if self.included_exog:
idx_columns_exog = np.arange(len(self.X_train_col_names))[len(self.lags) + step::self.steps]
idx_columns = np.hstack((idx_columns_lags, idx_columns_exog))
feature_names = [self.X_train_col_names[i] for i in idx_columns]
feature_names = [name.replace(f"_step_{step+1}", "") for name in feature_names]
try:
feature_importance = pd.DataFrame({
'feature': feature_names,
'importance' : estimator.feature_importances_
})
except:
try:
feature_importance = pd.DataFrame({
'feature': feature_names,
'importance' : estimator.coef_
})
except:
warnings.warn(
f"Impossible to access feature importance for regressor of type "
f"{type(estimator)}. This method is only valid when the "
f"regressor stores internally the feature importance in the "
f"attribute `feature_importances_` or `coef_`."
)
feature_importance = None
return feature_importance
predict(self, steps=None, last_window=None, exog=None)
¶
Predict n steps ahead.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
steps |
Union[int, list] |
Predict n steps. The value of If int: Only steps within the range of 1 to int are predicted. If list: List of ints. Only the steps contained in the list are predicted. If |
None |
last_window |
Optional[pandas.core.series.Series] |
Values of the series used to create the predictors (lags) need in the first iteration of prediction (t + 1). If |
None |
exog |
Union[pandas.core.series.Series, pandas.core.frame.DataFrame] |
Exogenous variable/s included as predictor/s. |
None |
Returns:
| Type | Description |
|---|---|
Series |
Predicted values. |
Source code in skforecast/ForecasterAutoregDirect/ForecasterAutoregDirect.py
def predict(
self,
steps: Optional[Union[int, list]]=None,
last_window: Optional[pd.Series]=None,
exog: Optional[Union[pd.Series, pd.DataFrame]]=None
) -> pd.Series:
"""
Predict n steps ahead.
Parameters
----------
steps : int, list, None, default `None`
Predict n steps. The value of `steps` must be less than or equal to the
value of steps defined when initializing the forecaster. Starts at 1.
If int:
Only steps within the range of 1 to int are predicted.
If list:
List of ints. Only the steps contained in the list are predicted.
If `None`:
As many steps are predicted as were defined at initialization.
last_window : pandas Series, default `None`
Values of the series used to create the predictors (lags) need in the
first iteration of 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.
Returns
-------
predictions : pandas Series
Predicted values.
"""
if isinstance(steps, int):
steps = list(range(steps))
elif steps is None:
steps = list(range(self.steps))
elif isinstance(steps, list):
steps = list(np.array(steps) - 1) # To start at 0 for indexing
for step in steps:
if not isinstance(step, (int, np.int64, np.int32)):
raise TypeError(
f"`steps` argument must be an int, a list of ints or `None`. "
f"Got {type(steps)}."
)
check_predict_input(
forecaster_type = type(self).__name__,
steps = steps,
fitted = self.fitted,
included_exog = self.included_exog,
index_type = self.index_type,
index_freq = self.index_freq,
window_size = self.window_size,
last_window = last_window,
exog = exog,
exog_type = self.exog_type,
exog_col_names = self.exog_col_names,
interval = None,
max_steps = self.steps,
levels = None,
series_col_names = None
)
if exog is not None:
if isinstance(exog, pd.DataFrame):
exog = transform_dataframe(
df = exog,
transformer = self.transformer_exog,
fit = False,
inverse_transform = False
)
else:
exog = transform_series(
series = exog,
transformer = self.transformer_exog,
fit = False,
inverse_transform = False
)
exog_values, _ = preprocess_exog(
exog = exog.iloc[:max(steps)+1, ]
)
exog_values = exog_to_direct(exog=exog_values, steps=max(steps)+1)
else:
exog_values = None
if last_window is None:
last_window = self.last_window.copy()
last_window = transform_series(
series = last_window,
transformer = self.transformer_y,
fit = False,
inverse_transform = False
)
last_window_values, last_window_index = preprocess_last_window(
last_window = last_window
)
X_lags = last_window_values[-self.lags].reshape(1, -1)
predictions = np.full(shape=len(steps), fill_value=np.nan)
for i, step in enumerate(steps):
regressor = self.regressors_[step]
if exog is None:
X = X_lags
else:
# Only columns from exog related with the current step are selected.
X = np.hstack([X_lags, exog_values[0][step::max(steps)+1].reshape(1, -1)])
with warnings.catch_warnings():
# Suppress scikit-learn warning: "X does not have valid feature names,
# but NoOpTransformer was fitted with feature names".
warnings.simplefilter("ignore")
predictions[i] = regressor.predict(X)
idx = expand_index(index=last_window_index, steps=max(steps)+1)
predictions = pd.Series(
data = predictions.reshape(-1),
index = idx[steps],
name = 'pred'
)
predictions = transform_series(
series = predictions,
transformer = self.transformer_y,
fit = False,
inverse_transform = True
)
return predictions
set_lags(self, lags)
¶
Set new value to the attribute lags.
Attributes max_lag and window_size are also updated.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
lags |
Union[int, list, numpy.ndarray, range] |
Lags used as predictors. Index starts at 1, so lag 1 is equal to t-1.
|
required |
Source code in skforecast/ForecasterAutoregDirect/ForecasterAutoregDirect.py
def set_lags(
self,
lags: Union[int, list, np.ndarray, range]
) -> None:
"""
Set new value to the attribute `lags`.
Attributes `max_lag` and `window_size` are also updated.
Parameters
----------
lags : int, list, 1D np.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`.
`list` or `np.ndarray`: include only lags present in `lags`.
Returns
-------
None
"""
self.lags = initialize_lags(type(self).__name__, lags)
self.max_lag = max(self.lags)
self.window_size = max(self.lags)
set_params(self, **params)
¶
Set new values to the parameters of the scikit learn model stored in the
forecaster. It is important to note that all models share the same configuration of parameters and hyperparameters.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
params |
dict |
Parameters values. |
{} |
Source code in skforecast/ForecasterAutoregDirect/ForecasterAutoregDirect.py
def set_params(
self,
**params: dict
) -> None:
"""
Set new values to the parameters of the scikit learn model stored in the
forecaster. It is important to note that all models share the same
configuration of parameters and hyperparameters.
Parameters
----------
params : dict
Parameters values.
Returns
-------
self
"""
self.regressor = clone(self.regressor)
self.regressor.set_params(**params)
self.regressors_ = {step: clone(self.regressor) for step in range(self.steps)}
_create_lags(self, y)
private
¶
Transforms a 1d array into a 2d array (X) and a 1d array (y). Each row
in X is associated with a value of y and it represents the lags that precede it.
Notice that, the returned matrix X_data, contains the lag 1 in the first column, the lag 2 in the second column and so on.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
y |
ndarray |
Training time series. |
required |
Returns:
| Type | Description |
|---|---|
Tuple[numpy.ndarray, numpy.ndarray] |
2d numpy array with the lagged values (predictors). |
Source code in skforecast/ForecasterAutoreg/ForecasterAutoreg.py
def _create_lags(
self,
y: np.ndarray
) -> Tuple[np.ndarray, np.ndarray]:
"""
Transforms a 1d array into a 2d array (X) and a 1d array (y). Each row
in X is associated with a value of y and it represents the lags that
precede it.
Notice that, the returned matrix X_data, contains the lag 1 in the first
column, the lag 2 in the second column and so on.
Parameters
----------
y : 1d numpy ndarray
Training time series.
Returns
-------
X_data : 2d numpy ndarray, shape (samples - max(self.lags), len(self.lags))
2d numpy array with the lagged values (predictors).
y_data : 1d numpy ndarray, shape (samples - max(self.lags),)
Values of the time series related to each row of `X_data`.
"""
n_splits = len(y) - self.max_lag
if n_splits <= 0:
raise ValueError(
f'The maximum lag ({self.max_lag}) must be less than the length '
f'of the series ({len(y)}).'
)
X_data = np.full(shape=(n_splits, len(self.lags)), fill_value=np.nan, dtype=float)
for i, lag in enumerate(self.lags):
X_data[:, i] = y[self.max_lag - lag: -lag]
y_data = y[self.max_lag:]
return X_data, y_data
__repr__(self)
special
¶
Information displayed when a ForecasterAutoreg object is printed.
Source code in skforecast/ForecasterAutoreg/ForecasterAutoreg.py
def __repr__(
self
) -> str:
"""
Information displayed when a ForecasterAutoreg object is printed.
"""
if isinstance(self.regressor, sklearn.pipeline.Pipeline):
name_pipe_steps = tuple(name + "__" for name in self.regressor.named_steps.keys())
params = {key : value for key, value in self.regressor.get_params().items() \
if key.startswith(name_pipe_steps)}
else:
params = self.regressor.get_params()
info = (
f"{'=' * len(str(type(self)).split('.')[1])} \n"
f"{str(type(self)).split('.')[1]} \n"
f"{'=' * len(str(type(self)).split('.')[1])} \n"
f"Regressor: {self.regressor} \n"
f"Lags: {self.lags} \n"
f"Transformer for y: {self.transformer_y} \n"
f"Transformer for exog: {self.transformer_exog} \n"
f"Window size: {self.window_size} \n"
f"Weight function included: {True if self.weight_func is not None else False} \n"
f"Exogenous included: {self.included_exog} \n"
f"Type of exogenous variable: {self.exog_type} \n"
f"Exogenous variables names: {self.exog_col_names} \n"
f"Training range: {self.training_range.to_list() if self.fitted else None} \n"
f"Training index type: {str(self.index_type).split('.')[-1][:-2] if self.fitted else None} \n"
f"Training index frequency: {self.index_freq if self.fitted else None} \n"
f"Regressor parameters: {params} \n"
f"Creation date: {self.creation_date} \n"
f"Last fit date: {self.fit_date} \n"
f"Skforecast version: {self.skforcast_version} \n"
f"Python version: {self.python_version} \n"
)
return info