make_reduction#

make_reduction(estimator, strategy='recursive', window_length=10, scitype='infer', transformers=None, pooling='local', windows_identical=True)[source]#

Make forecaster based on reduction to tabular or time-series regression.

During fitting, a sliding-window approach is used to first transform the time series into tabular or panel data, which is then used to fit a tabular or time-series regression estimator. During prediction, the last available data is used as input to the fitted regression estimator to generate forecasts.

Please see below a graphical representation of the make_reduction logic using the following symbols:

y = forecast target.
x = past values of y that are used as features (X) to forecast y
* = observations, past or future, neither part of window nor forecast.

Assume we have the following training data (14 observations):

|----------------------------|
| * * * * * * * * * * * * * *|
|----------------------------|

And want to forecast with window_length = 9 and fh = [2, 4].

By construction, a recursive reducer always targets the first data point after the window, irrespective of the forecasting horizons requested. In the example the following 5 windows are created:

|----------------------------|
| x x x x x x x x x y * * * *|
| * x x x x x x x x x y * * *|
| * * x x x x x x x x x y * *|
| * * * x x x x x x x x x y *|
| * * * * x x x x x x x x x y|
|----------------------------|

Direct Reducers will create multiple models, one for each forecasting horizon. With the argument windows_identical = True (default) the windows used to train the model are defined by the maximum forecasting horizon. Only two complete windows can be defined in this example fh = 4 (maximum of fh = [2, 4]):

|----------------------------|
| x x x x x x x x x * * * y *|
| * x x x x x x x x x * * * y|
|----------------------------|

All other forecasting horizons will also use those two (maximal) windows. fh = 2:

|----------------------------|
| x x x x x x x x x * y * * *|
| * x x x x x x x x x * y * *|
|----------------------------|

With windows_identical = False we drop the requirement to use the same windows for each of the direct models, so more windows can be created for horizons other than the maximum forecasting horizon. fh = 2:

|----------------------------|
| x x x x x x x x x * y * * *|
| * x x x x x x x x x * y * *|
| * * x x x x x x x x x * y *|
| * * * x x x x x x x x x * y|
|----------------------------|

fh = 4:

|----------------------------|
| x x x x x x x x x * * * y *|
| * x x x x x x x x x * * * y|
|----------------------------|

Use windows_identical = True if you want to compare the forecasting performance across different horizons, since all models trained will use the same windows. Use windows_identical = False if you want to have the highest forecasting accuracy for each forecasting horizon.

Parameters:

estimatoran estimator instance, can be:

scikit-learn regressor or interface compatible
sktime time series regressor
skpro tabular probabilistic supervised regressor, only for direct reduction this will result in a probabilistic forecaster

strategystr, optional (default=”recursive”)

The strategy to generate forecasts. Must be one of “direct”, “recursive” or “multioutput”.

window_lengthint, optional (default=10)

Window length used in sliding window transformation.

scitypestr, optional (default=”infer”)

Legacy argument for downwards compatibility, should not be used. make_reduction will automatically infer the correct type of estimator. This internal inference can be force-overridden by the scitype argument. Must be one of “infer”, “tabular-regressor” or “time-series-regressor”. If the scitype cannot be inferred, this is a bug and should be reported.

transformers: list of transformers (default = None)

A suitable list of transformers that allows for using an en-bloc approach with make_reduction. This means that instead of using the raw past observations of y across the window length, suitable features will be generated directly from the past raw observations. Currently only supports WindowSummarizer (or a list of WindowSummarizers) to generate features e.g. the mean of the past 7 observations. Currently only works for RecursiveTimeSeriesRegressionForecaster.

pooling: str {“local”, “global”}, optional

Specifies whether separate models will be fit at the level of each instance (local) of if you wish to fit a single model to all instances (“global”). Currently only works for RecursiveTimeSeriesRegressionForecaster.

windows_identical: bool, (default = True)

Direct forecasting only. Specifies whether all direct models use the same X windows from y (True: Number of windows = total observations + 1 - window_length - maximum forecasting horizon) or a different number of X windows depending on the forecasting horizon (False: Number of windows = total observations + 1 - window_length - forecasting horizon). See pictionary below for more information.

Returns:

forecasteran sktime forecaster object: the reduction forecaster, wrapping estimator class is determined by the strategy argument and type of estimator.

References

[1]

Bontempi, Gianluca & Ben Taieb, Souhaib & Le Borgne, Yann-Aël. (2013). Machine Learning Strategies for Time Series Forecasting.

Examples

>>> from sktime.forecasting.compose import make_reduction
>>> from sktime.datasets import load_airline
>>> from sklearn.ensemble import GradientBoostingRegressor
>>> y = load_airline()
>>> regressor = GradientBoostingRegressor()
>>> forecaster = make_reduction(regressor, window_length=15, strategy="recursive")
>>> forecaster.fit(y)
RecursiveTabularRegressionForecaster(...)
>>> y_pred = forecaster.predict(fh=[1,2,3])