# !/usr/bin/env python3 -u
# copyright: sktime developers, BSD-3-Clause License (see LICENSE file)
"""Implements simple forecasts based on naive assumptions."""
__all__ = ["NaiveForecaster", "NaiveVariance"]
__author__ = [
"mloning",
"piyush1729",
"sri1419",
"Flix6x",
"aiwalter",
"IlyasMoutawwakil",
"fkiraly",
"bethrice44",
]
import math
import numpy as np
import pandas as pd
from scipy.stats import norm
from sktime.datatypes._convert import convert, convert_to
from sktime.datatypes._utilities import get_slice
from sktime.forecasting.base import ForecastingHorizon
from sktime.forecasting.base._base import DEFAULT_ALPHA, BaseForecaster
from sktime.forecasting.base._sktime import _BaseWindowForecaster
from sktime.utils.seasonality import _pivot_sp, _unpivot_sp
from sktime.utils.validation import check_window_length
from sktime.utils.validation.forecasting import check_sp
from sktime.utils.warnings import warn
[文档]class NaiveForecaster(_BaseWindowForecaster):
"""Forecast based on naive assumptions about past trends continuing.
NaiveForecaster is a forecaster that makes forecasts using simple
strategies. Two out of three strategies are robust against NaNs. The
NaiveForecaster can also be used for multivariate data and it then
applies internally the ColumnEnsembleForecaster, so each column
is forecasted with the same strategy.
Internally, this forecaster does the following:
- obtains the so-called "last window", a 1D array that denotes the
most recent time window that the forecaster is allowed to use
- reshapes the last window into a 2D array according to the given
seasonal periodicity (prepended with NaN values to make it fit);
- make a prediction for each column, using the given strategy:
- "last": last non-NaN row
- "mean": np.nanmean over rows
- tile the predictions using the seasonal periodicity
To compute prediction quantiles, we first estimate the standard error
of prediction residuals under the assumption of uncorrelated residuals.
The forecast variance is then computed by multiplying the residual
variance by a constant. This constant is a small-sample bias adjustment
and each method (mean, last, drift) have different formulas for computing
the constant. These formulas can be found in the Forecasting:
Principles and Practice textbook (Table 5.2) [1]_. Lastly, under the assumption that
residuals follow a normal distribution, we use the forecast variance and
z-scores of a normal distribution to estimate the prediction quantiles.
Parameters
----------
strategy : {"last", "mean", "drift"}, default="last"
Strategy used to make forecasts:
* "last": (robust against NaN values)
forecast the last value in the
training series when sp is 1.
When sp is not 1,
last value of each season
in the last window will be
forecasted for each season.
* "mean": (robust against NaN values)
forecast the mean of last window
of training series when sp is 1.
When sp is not 1, mean of all values
in a season from last window will be
forecasted for each season.
* "drift": (not robust against NaN values)
forecast by fitting a line between the
first and last point of the window and
extrapolating it into the future.
sp : int, or None, default=1
Seasonal periodicity to use in the seasonal forecasting. None=1.
window_length : int or None, default=None
Window length to use in the ``mean`` strategy. If None, entire training
series will be used.
References
----------
.. [1] Hyndman, R.J., & Athanasopoulos, G. (2021) Forecasting:
principles and practice, 3rd edition, OTexts: Melbourne, Australia.
OTexts.com/fpp3. Accessed on 22 September 2022.
Examples
--------
>>> from sktime.datasets import load_airline
>>> from sktime.forecasting.naive import NaiveForecaster
>>> y = load_airline()
>>> forecaster = NaiveForecaster(strategy="drift")
>>> forecaster.fit(y)
NaiveForecaster(...)
>>> y_pred = forecaster.predict(fh=[1,2,3])
"""
_tags = {
# packaging info
# --------------
"authors": [
"mloning",
"piyush1729",
"sri1419",
"Flix6x",
"aiwalter",
"IlyasMoutawwakil",
"fkiraly",
"bethrice44",
],
# estimator type
# --------------
"y_inner_mtype": "pd.Series",
"requires-fh-in-fit": False,
"handles-missing-data": True,
"ignores-exogeneous-X": True,
"scitype:y": "univariate",
"capability:pred_var": True,
"capability:pred_int": True,
}
def __init__(self, strategy="last", window_length=None, sp=1):
super().__init__()
self.strategy = strategy
self.sp = sp
self.window_length = window_length
# Override tag for handling missing data
# todo: remove if GH1367 is fixed
if self.strategy in ("last", "mean"):
self.set_tags(**{"handles-missing-data": True})
def _fit(self, y, X, fh):
"""Fit to training data.
Parameters
----------
y : pd.Series
Target time series to which to fit the forecaster.
fh : int, list or np.array, default=None
The forecasters horizon with the steps ahead to to predict.
X : pd.DataFrame, default=None
Exogenous variables are ignored.
Returns
-------
self : returns an instance of self.
"""
# X_train is ignored
sp = self.sp or 1
n_timepoints = y.shape[0]
if self.strategy in ("last", "mean"):
# check window length is greater than sp for seasonal mean or seasonal last
if self.window_length is not None and sp != 1:
if self.window_length < sp:
raise ValueError(
f"The `window_length`: "
f"{self.window_length} is smaller than "
f"`sp`: {sp}."
)
self.window_length_ = check_window_length(self.window_length, n_timepoints)
self.sp_ = check_sp(sp)
# if not given, set default window length
if self.window_length is None:
self.window_length_ = len(y)
elif self.strategy == "drift":
if sp != 1:
warn(
"For the `drift` strategy, the `sp` value will be ignored.",
obj=self,
)
# window length we need for forecasts is just the
# length of seasonal periodicity
self.window_length_ = check_window_length(self.window_length, n_timepoints)
if self.window_length is None:
self.window_length_ = len(y)
if self.window_length == 1:
raise ValueError(
f"For the `drift` strategy, "
f"the `window_length`: {self.window_length} "
f"value must be greater than one."
)
else:
allowed_strategies = ("last", "mean", "drift")
raise ValueError(
f"Unknown strategy: {self.strategy}. Expected "
f"one of: {allowed_strategies}."
)
# check window length
if self.window_length_ > len(y):
param = "sp" if self.strategy == "last" and sp != 1 else "window_length_"
raise ValueError(
f"The {param}: {self.window_length_} is larger than "
f"the training series."
)
return self
def _predict_last_window(
self, fh, X=None, return_pred_int=False, alpha=DEFAULT_ALPHA
):
"""Calculate predictions for use in predict."""
last_window, _ = self._get_last_window()
fh = fh.to_relative(self.cutoff)
strategy = self.strategy
sp = self.sp or 1
# if last window only contains missing values, return nan
if np.all(np.isnan(last_window)) or len(last_window) == 0:
return self._predict_nan(fh)
elif strategy == "last" or (strategy == "drift" and self.window_length_ == 1):
if sp == 1:
last_valid_value = last_window[
(~np.isnan(last_window))[0::sp].cumsum().argmax()
]
return np.repeat(last_valid_value, len(fh))
else:
# reshape last window, one column per season
last_window = self._reshape_last_window_for_sp(last_window)
# select last non-NaN row for every column
y_pred = last_window[
(~np.isnan(last_window)).cumsum(0).argmax(0).T,
range(last_window.shape[1]),
]
# tile prediction according to seasonal periodicity
y_pred = self._tile_seasonal_prediction(y_pred, fh)
elif strategy == "mean":
if sp == 1:
return np.repeat(np.nanmean(last_window), len(fh))
else:
# reshape last window, one column per season
last_window = self._reshape_last_window_for_sp(last_window)
# compute seasonal mean, averaging over non-NaN rows for every column
y_pred = np.nanmean(last_window, axis=0)
# tile prediction according to seasonal periodicity
y_pred = self._tile_seasonal_prediction(y_pred, fh)
elif strategy == "drift":
if self.window_length_ != 1:
if np.any(np.isnan(last_window[[0, -1]])):
raise ValueError(
f"For {strategy},"
f"first and last elements in the last "
f"window must not be a missing value."
)
else:
# formula for slope
slope = (last_window[-1] - last_window[0]) / (
self.window_length_ - 1
)
# get zero-based index by subtracting the minimum
fh_idx = fh.to_indexer(self.cutoff)
# linear extrapolation
y_pred = last_window[-1] + (fh_idx + 1) * slope
else:
raise ValueError(f"unknown strategy {strategy} provided to NaiveForecaster")
return y_pred
def _reshape_last_window_for_sp(self, last_window):
"""Reshape the 1D last window into a 2D last window, prepended with NaN values.
The 2D array has 1 column per season.
For example:
last_window = [1, 2, 3, 4]
sp = 3 # i.e. 3 distinct seasons
reshaped_last_window = [[nan, nan, 1],
[ 2, 3, 4]]
"""
# if window length is not multiple of sp, backward fill window with nan values
remainder = self.window_length_ % self.sp_
if remainder > 0:
pad_width = self.sp_ - remainder
else:
pad_width = 0
pad_width += self.window_length_ - len(last_window)
last_window = np.hstack([np.full(pad_width, np.nan), last_window])
# reshape last window, one column per season
last_window = last_window.reshape(
int(np.ceil(self.window_length_ / self.sp_)), self.sp_
)
return last_window
def _tile_seasonal_prediction(self, y_pred, fh):
"""Tile a prediction to cover all requested forecasting horizons.
The original prediction has 1 value per season.
For example:
fh = [1, 2, 3, 4, 5, 6, 7]
y_pred = [2, 3, 1] # note len(y_pred) = sp
y_pred_tiled = [2, 3, 1, 2, 3, 1, 2]
"""
# we need to replicate the last window if max(fh) is
# larger than sp,
# so that we still make forecasts by repeating the
# last value for that season,
# assume fh is sorted, i.e. max(fh) == fh[-1]
# only slicing all the last seasons into last_window
if fh[-1] > self.sp_:
reps = int(np.ceil(fh[-1] / self.sp_))
y_pred = np.tile(y_pred, reps=reps)
# get zero-based index by subtracting the minimum
fh_idx = fh.to_indexer(self.cutoff)
return y_pred[fh_idx]
def _predict_naive(self, fh=None, X=None):
from sktime.transformations.series.lag import Lag
strategy = self.strategy
sp = self.sp
_y = self._y
cutoff = self.cutoff
if isinstance(_y.index, pd.DatetimeIndex) and hasattr(_y.index, "freq"):
freq = _y.index.freq
else:
freq = None
lagger = Lag(1, keep_column_names=True, freq=freq)
expected_index = fh.to_absolute(cutoff).to_pandas()
if strategy == "last" and sp == 1:
y_old = lagger.fit_transform(_y)
y_new = pd.DataFrame(index=expected_index, columns=[0], dtype="float64")
full_y = pd.concat([y_old, y_new], keys=["a", "b"]).sort_index(level=-1)
y_filled = full_y.ffill().bfill()
# subset to rows that contain elements we wanted to fill
y_pred = y_filled.loc["b"]
# convert to pd.Series from pd.DataFrame
y_pred = y_pred.iloc[:, 0]
elif strategy == "last" and sp > 1:
y_old = _pivot_sp(_y, sp, anchor_side="end")
y_old = lagger.fit_transform(y_old)
y_new_mask = pd.Series(index=expected_index, dtype="float64")
y_new = _pivot_sp(y_new_mask, sp, anchor=_y, anchor_side="end")
full_y = pd.concat([y_old, y_new], keys=["a", "b"]).sort_index(level=-1)
y_filled = full_y.ffill().bfill()
# subset to rows that contain elements we wanted to fill
y_pred = y_filled.loc["b"]
# reformat to wide
y_pred = _unpivot_sp(y_pred, template=_y)
# subset to required indices
y_pred = y_pred.reindex(expected_index)
# convert to pd.Series from pd.DataFrame
y_pred = y_pred.iloc[:, 0]
y_pred.name = _y.name
return y_pred
def _predict(self, fh=None, X=None):
"""Forecast time series at future horizon.
Parameters
----------
fh : int, list, np.array or ForecastingHorizon
Forecasting horizon
X : pd.DataFrame, optional (default=None)
Exogenous time series
"""
strategy = self.strategy
NEW_PREDICT = ["last"]
if strategy in NEW_PREDICT:
return self._predict_naive(fh=fh, X=X)
y_pred = super()._predict(fh=fh, X=X)
# test_predict_time_index_in_sample_full[ForecastingPipeline-0-int-int-True]
# causes a pd.DataFrame to appear as y_pred, which upsets the next lines
# reasons are unclear, this is coming from the _BaseWindowForecaster
# todo: investigate this
if isinstance(y_pred, pd.DataFrame):
y_pred = y_pred.iloc[:, 0]
# check for in-sample prediction, if first time point needs to be imputed
if self._y.index[0] in y_pred.index:
if y_pred.loc[[self._y.index[0]]].hasnans:
# fill NaN with observed values
y_pred.loc[self._y.index[0]] = self._y[self._y.index[1]]
y_pred.name = self._y.name
return y_pred
def _predict_quantiles(self, fh, X, alpha):
"""Compute/return prediction quantiles for a forecast.
Uses normal distribution as predictive distribution to compute the
quantiles.
Parameters
----------
fh : int, list, np.array or ForecastingHorizon
Forecasting horizon
X : pd.DataFrame, optional (default=None)
Exogenous variables are ignored.
alpha : float or list of float, optional (default=0.5)
A probability or list of, at which quantile forecasts are computed.
Returns
-------
pred_quantiles : pd.DataFrame
Column has multi-index: first level is variable name from y in fit,
second level being the quantile forecasts for each alpha.
Quantile forecasts are calculated for each a in alpha.
Row index is fh. Entries are quantile forecasts, for var in col index,
at quantile probability in second-level col index, for each row index.
"""
y_pred = self.predict(fh)
y_pred = convert(y_pred, from_type=self._y_mtype_last_seen, to_type="pd.Series")
pred_var = self.predict_var(fh)
z_scores = norm.ppf(alpha)
errors = (
np.sqrt(pred_var.to_numpy().reshape(len(pred_var), 1)) * z_scores
).reshape(len(y_pred), len(alpha))
var_names = self._get_varnames()
pred_quantiles = pd.DataFrame(
errors + y_pred.values.reshape(len(y_pred), 1),
columns=pd.MultiIndex.from_product([var_names, alpha]),
index=fh.to_absolute_index(self.cutoff),
)
return pred_quantiles
def _predict_var(self, fh, X=None, cov=False):
"""Compute/return prediction variance for naive forecasts.
Variance are computed according to formulas from (Table 5.2)
Forecasting: Principles and Practice textbook [1]_.
Must be run *after* the forecaster has been fitted.
Parameters
----------
fh : int, list, np.array or ForecastingHorizon
Forecasting horizon
X : pd.DataFrame, optional (default=None)
Exogenous variables are ignored.
cov : bool, optional (default=False)
If True, return the covariance matrix.
If False, return the marginal variance.
Returns
-------
pred_var :
if cov=False, pd.DataFrame with index fh.
a vector of same length as fh with predictive marginal variances;
if cov=True, pd.DataFrame with index fh and columns fh.
a square matrix of size len(fh) with predictive covariance matrix.
References
----------
.. [1] https://otexts.com/fpp3/prediction-intervals.html#benchmark-methods
"""
y = self._y
y = convert_to(y, "pd.Series")
T = len(y)
sp = self.sp
# Compute "past" residuals
if self.strategy == "last":
y_res = y - y.shift(self.sp)
elif self.strategy == "mean":
if not self.window_length:
if sp > 1:
# Get the next sp predictions and repeat them
# T / self.sp times to match the length of trained y
# NOTE: +1 extra tile to defend against off-by-one errors
reps = math.ceil(T / sp) + 1
past_fh = ForecastingHorizon(
list(range(1, sp + 1)), is_relative=None, freq=self._cutoff
)
seasonal_means = self._predict(fh=past_fh)
if isinstance(seasonal_means, pd.DataFrame):
seasonal_means = seasonal_means.squeeze()
y_pred = np.tile(seasonal_means.to_numpy(), reps)[0:T]
else:
# Since this strategy returns a constant, just predict fh=1 and
# transform the constant into a repeated array
past_fh = ForecastingHorizon(1, is_relative=None, freq=self._cutoff)
y_pred = np.repeat(np.squeeze(self._predict(fh=past_fh)), T)
else:
if sp > 1:
# Label index by seasonal period
seasons = np.mod(np.arange(T), sp)
# Compute rolling seasonal means
y_pred = (
y.to_frame()
.assign(__sp__=seasons)
# Group observations by their
# seasonal period position
.groupby("__sp__")
# Compute rolling means per
# seasonal period
.rolling(self.window_length)
.mean()
.droplevel("__sp__")
.sort_index()
.squeeze()
)
else:
# Compute rolling means
y_pred = y.rolling(self.window_length).mean()
y_res = y - y_pred
else:
# Slope equation from:
# https://otexts.com/fpp3/simple-methods.html#drift-method
slope = (y.iloc[-1] - y.iloc[-(self.window_length or 0)]) / (T - 1)
# Fitted value = previous value + slope
# https://github.com/robjhyndman/forecast/blob/master/R/naive.R#L34
y_res = y - (y.shift(sp) + slope)
# Residuals MSE and SE
# + 1 degrees of freedom to estimate drift coefficient standard error
# https://github.com/robjhyndman/forecast/blob/master/R/naive.R#L79
n_nans = np.sum(pd.isna(y_res))
mse_res = np.sum(np.square(y_res)) / (T - n_nans - (self.strategy == "drift"))
se_res = np.sqrt(mse_res)
window_length = self.window_length or T
def sqrt_flr(x):
"""Square root of x, floored at 1 - to deal with in-sample predictions."""
return np.sqrt(np.maximum(x, 1))
# Formulas from:
# https://otexts.com/fpp3/prediction-intervals.html (Table 5.2)
partial_se_formulas = {
"last": (
sqrt_flr if sp == 1 else lambda h: sqrt_flr(np.floor((h - 1) / sp) + 1)
),
"mean": lambda h: np.repeat(sqrt_flr(1 + (1 / window_length)), len(h)),
"drift": lambda h: sqrt_flr(h * (1 + (h / (T - 1)))),
}
fh_periods = np.array(fh.to_relative(self.cutoff))
marginal_se = se_res * partial_se_formulas[self.strategy](fh_periods)
marginal_vars = marginal_se**2
fh_idx = fh.to_absolute_index(self.cutoff)
if cov:
fh_size = len(fh)
cov_matrix = np.fill_diagonal(
np.zeros(shape=(fh_size, fh_size)), marginal_vars
)
pred_var = pd.DataFrame(cov_matrix, columns=fh_idx, index=fh_idx)
else:
pred_var = pd.DataFrame(marginal_vars, index=fh_idx)
return pred_var
[文档] @classmethod
def get_test_params(cls, parameter_set="default"):
"""Return testing parameter settings for the estimator.
Parameters
----------
parameter_set : str, default="default"
Name of the set of test parameters to return, for use in tests. If no
special parameters are defined for a value, will return ``"default"`` set.
Returns
-------
params : dict or list of dict, default = {}
Parameters to create testing instances of the class
Each dict are parameters to construct an "interesting" test instance, i.e.,
``MyClass(**params)`` or ``MyClass(**params[i])`` creates a valid test
instance.
``create_test_instance`` uses the first (or only) dictionary in ``params``
"""
params_list = [
{},
{"strategy": "mean", "sp": 2},
{"strategy": "drift"},
{"strategy": "last"},
{"strategy": "mean", "window_length": 5},
]
return params_list
[文档]class NaiveVariance(BaseForecaster):
r"""Compute the prediction variance based on a naive strategy.
NaiveVariance adds to a ``forecaster`` the ability to compute the
prediction variance based on naive assumptions about the time series.
The simple strategy is as follows:
- Let :math:`y_1,\dots,y_T` be the time series we fit the estimator :math:`f` to.
- Let :math:`\widehat{y}_{ij}` be the forecast for time point :math:`j`, obtained
from fitting the forecaster to the partial time series :math:`y_1,\dots,y_i`.
- We compute the residuals matrix :math:`R=(r_{ij})=(y_j-\widehat{y}_{ij})`.
- The variance prediction :math:`v_k` for :math:`y_{T+k}` is
:math:`\frac{1}{T-k}\sum_{i=1}^{T-k} a_{i,i+k}^2`
because we are averaging squared residuals of all forecasts that are :math:`k`
time points ahead.
- And for the covariance matrix prediction, the formula becomes
:math:`Cov(y_k, y_l)=\frac{\sum_{i=1}^N \hat{r}_{k,k+i}*\hat{r}_{l,l+i}}{N}`.
The resulting forecaster will implement
``predict_interval``, ``predict_quantiles``, ``predict_var``, and
``predict_proba``,
even if the wrapped forecaster ``forecaster`` did not have this capability;
for point forecasts (``predict``), behaves like the wrapped forecaster.
Parameters
----------
forecaster : estimator
Estimator to which probabilistic forecasts are being added
initial_window : int, optional, default=1
number of minimum initial indices to use for fitting when computing residuals
verbose : bool, optional, default=False
whether to print warnings if windows with too few data points occur
Examples
--------
>>> from sktime.datasets import load_airline
>>> from sktime.forecasting.naive import NaiveForecaster, NaiveVariance
>>> y = load_airline()
>>> forecaster = NaiveForecaster(strategy="drift")
>>> variance_forecaster = NaiveVariance(forecaster)
>>> variance_forecaster.fit(y)
NaiveVariance(...)
>>> var_pred = variance_forecaster.predict_var(fh=[1,2,3])
"""
_tags = {
# packaging info
# --------------
"authors": ["fkiraly", "bethrice44"],
# estimator type
# --------------
"scitype:y": "univariate",
"requires-fh-in-fit": False,
"handles-missing-data": False,
"ignores-exogeneous-X": False,
"capability:pred_int": True,
"capability:pred_var": True,
}
def __init__(self, forecaster, initial_window=1, verbose=False):
self.forecaster = forecaster
self.initial_window = initial_window
self.verbose = verbose
super().__init__()
tags_to_clone = [
"requires-fh-in-fit",
"ignores-exogeneous-X",
"handles-missing-data",
"y_inner_mtype",
"X_inner_mtype",
"X-y-must-have-same-index",
"enforce_index_type",
]
self.clone_tags(self.forecaster, tags_to_clone)
def _fit(self, y, X, fh):
self.fh_early_ = fh is not None
self.forecaster_ = self.forecaster.clone()
self.forecaster_.fit(y=y, X=X, fh=fh)
if self.fh_early_:
self.residuals_matrix_ = self._compute_sliding_residuals(
y=y, X=X, forecaster=self.forecaster, initial_window=self.initial_window
)
return self
def _predict(self, fh, X):
return self.forecaster_.predict(fh=fh, X=X)
def _update(self, y, X=None, update_params=True):
self.forecaster_.update(y, X, update_params=update_params)
if update_params and self._fh is not None:
self.residuals_matrix_ = self._compute_sliding_residuals(
y=self._y,
X=self._X,
forecaster=self.forecaster,
initial_window=self.initial_window,
)
return self
def _predict_quantiles(self, fh, X, alpha):
"""Compute/return prediction quantiles for a forecast.
Uses normal distribution as predictive distribution to compute the
quantiles.
Parameters
----------
fh : int, list, np.array or ForecastingHorizon
Forecasting horizon
X : pd.DataFrame, optional (default=None)
Exogenous time series
alpha : float or list of float, optional (default=0.5)
A probability or list of, at which quantile forecasts are computed.
Returns
-------
pred_quantiles : pd.DataFrame
Column has multi-index: first level is variable name from y in fit,
second level being the quantile forecasts for each alpha.
Quantile forecasts are calculated for each a in alpha.
Row index is fh. Entries are quantile forecasts, for var in col index,
at quantile probability in second-level col index, for each row index.
"""
y_pred = self.predict(fh, X)
y_pred = convert(y_pred, from_type=self._y_mtype_last_seen, to_type="pd.Series")
pred_var = self.predict_var(fh, X)
pred_var = pred_var[pred_var.columns[0]]
pred_var.index = y_pred.index
z_scores = norm.ppf(alpha)
errors = [pred_var**0.5 * z for z in z_scores]
var_names = self._get_varnames()
var_name = var_names[0]
index = pd.MultiIndex.from_product([var_names, alpha])
pred_quantiles = pd.DataFrame(columns=index)
for a, error in zip(alpha, errors):
pred_quantiles[(var_name, a)] = y_pred + error
fh_absolute = fh.to_absolute(self.cutoff)
pred_quantiles.index = fh_absolute.to_pandas()
return pred_quantiles
def _predict_var(self, fh, X=None, cov=False):
"""Compute/return prediction variance for a forecast.
Must be run *after* the forecaster has been fitted.
Parameters
----------
fh : int, list, np.array or ForecastingHorizon
Forecasting horizon
X : pd.DataFrame, optional (default=None)
Exogenous time series
cov : bool, optional (default=False)
If True, return the covariance matrix.
If False, return the marginal variance.
Returns
-------
pred_var :
if cov=False, pd.DataFrame with index fh.
a vector of same length as fh with predictive marginal variances;
if cov=True, pd.DataFrame with index fh and columns fh.
a square matrix of size len(fh) with predictive covariance matrix.
"""
if self.fh_early_:
residuals_matrix = self.residuals_matrix_
else:
residuals_matrix = self._compute_sliding_residuals(
y=self._y,
X=self._X,
forecaster=self.forecaster,
initial_window=self.initial_window,
)
fh_relative = fh.to_relative(self.cutoff)
fh_absolute = fh.to_absolute(self.cutoff)
fh_absolute_ix = fh_absolute.to_pandas()
if cov:
fh_size = len(fh)
covariance = np.zeros(shape=(len(fh), fh_size))
for i in range(fh_size):
i_residuals = np.diagonal(residuals_matrix, offset=fh_relative[i])
for j in range(i, fh_size): # since the matrix is symmetric
j_residuals = np.diagonal(residuals_matrix, offset=fh_relative[j])
max_residuals = min(len(j_residuals), len(i_residuals))
covariance[i, j] = covariance[j, i] = np.nanmean(
i_residuals[:max_residuals] * j_residuals[:max_residuals]
)
pred_var = pd.DataFrame(
covariance,
index=fh_absolute_ix,
columns=fh_absolute_ix,
)
else:
variance = [
np.nanmean(np.diagonal(residuals_matrix, offset=offset) ** 2)
for offset in fh_relative
]
if hasattr(self._y, "columns"):
columns = self._y.columns
pred_var = pd.DataFrame(variance, columns=columns, index=fh_absolute_ix)
else:
pred_var = pd.DataFrame(variance, index=fh_absolute_ix)
return pred_var
def _compute_sliding_residuals(self, y, X, forecaster, initial_window):
"""Compute sliding residuals used in uncertainty estimates.
Parameters
----------
y : pd.Series or pd.DataFrame
sktime compatible time series to use in computing residuals matrix
X : pd.DataFrame
sktime compatible exogeneous time series to use in forecasts
forecaster : sktime compatible forecaster
forecaster to use in computing the sliding residuals
initial_window : int
minimum length of initial window to use in fitting
Returns
-------
residuals_matrix : pd.DataFrame, row and column index = y.index[initial_window:]
[i,j]-th entry is signed residual of forecasting y.loc[j] from y.loc[:i],
using a clone of the forecaster passed through the forecaster arg
"""
y = convert_to(y, "pd.Series")
y_index = y.index[initial_window:]
residuals_matrix = pd.DataFrame(columns=y_index, index=y_index, dtype="float")
for id in y_index:
forecaster = forecaster.clone()
y_train = get_slice(y, start=None, end=id) # subset on which we fit
y_test = get_slice(y, start=id, end=None) # subset on which we predict
try:
forecaster.fit(y_train, fh=y_test.index)
except ValueError:
if self.verbose:
warn(
f"Couldn't fit the model on "
f"time series window length {len(y_train)}.\n",
obj=self,
)
continue
try:
residuals_matrix.loc[id] = forecaster.predict_residuals(y_test, X)
except IndexError:
warn(
f"Couldn't predict after fitting on time series of length "
f"{len(y_train)}.\n",
obj=self,
)
return residuals_matrix
[文档] @classmethod
def get_test_params(cls, parameter_set="default"):
"""Return testing parameter settings for the estimator.
Parameters
----------
parameter_set : str, default="default"
Name of the set of test parameters to return, for use in tests. If no
special parameters are defined for a value, will return ``"default"`` set.
Returns
-------
params : dict or list of dict
"""
from sktime.forecasting.naive import NaiveForecaster
FORECASTER = NaiveForecaster()
params1 = {"forecaster": FORECASTER}
params2 = {"forecaster": FORECASTER, "initial_window": 2}
return [params1, params2]
