Transformer Selection

Load Data

from sktime.datasets import load_airline
from sktime.forecasting.model_selection import temporal_train_test_split

y = load_airline() # 144 for 12 years
y_train, y_test = temporal_train_test_split(y, test_size=36) # hold out last 3 years

from sklearn.preprocessing import StandardScaler

from sktime.forecasting.compose import TransformedTargetForecaster
from sktime.forecasting.model_selection import ForecastingGridSearchCV, SlidingWindowSplitter
from sktime.forecasting.naive import NaiveForecaster
from sktime.transformations.compose import OptionalPassthrough
from sktime.transformations.series.adapt import TabularToSeriesAdaptor
from sktime.transformations.series.detrend import Deseasonalizer

Create Pipeline

# decide if use a transformer with OptionalPassthrough
# passthrough, False, use the transformer within, True, ignores the transformer within
pipe = TransformedTargetForecaster(
    steps=[
        ("deseasonalizer", OptionalPassthrough(Deseasonalizer())),
        ("scaler", OptionalPassthrough(TabularToSeriesAdaptor(StandardScaler()))),
        ("forecaster", NaiveForecaster()),
    ]
)

pipe.get_params()

{'steps': [('deseasonalizer',
   OptionalPassthrough(transformer=Deseasonalizer())),
  ('scaler',
   OptionalPassthrough(transformer=TabularToSeriesAdaptor(transformer=StandardScaler()))),
  ('forecaster', NaiveForecaster())],
 'deseasonalizer': OptionalPassthrough(transformer=Deseasonalizer()),
 'scaler': OptionalPassthrough(transformer=TabularToSeriesAdaptor(transformer=StandardScaler())),
 'forecaster': NaiveForecaster(),
 'deseasonalizer__passthrough': False,
 'deseasonalizer__transformer__model': 'additive',
 'deseasonalizer__transformer__sp': 1,
 'deseasonalizer__transformer': Deseasonalizer(),
 'scaler__passthrough': False,
 'scaler__transformer__fit_in_transform': False,
 'scaler__transformer__transformer__copy': True,
 'scaler__transformer__transformer__with_mean': True,
 'scaler__transformer__transformer__with_std': True,
 'scaler__transformer__transformer': StandardScaler(),
 'scaler__transformer': TabularToSeriesAdaptor(transformer=StandardScaler()),
 'forecaster__sp': 1,
 'forecaster__strategy': 'last',
 'forecaster__window_length': None}

Select Transformer

cv = SlidingWindowSplitter(initial_window=60, window_length=24, start_with_window=True, step_length=24)

param_grid = {
    "deseasonalizer__passthrough": [True, False],
    "scaler__transformer__transformer__with_mean": [True, False],
    "scaler__passthrough": [True, False],
    "forecaster__strategy": ["drift", "mean", "last"],
}

gscv = ForecastingGridSearchCV(forecaster=pipe, param_grid=param_grid, cv=cv, n_jobs=-1)

selection = gscv.fit(y_train)

selection.best_params_

{'deseasonalizer__passthrough': True,
 'forecaster__strategy': 'drift',
 'scaler__passthrough': True,
 'scaler__transformer__transformer__with_mean': True}

model = selection.best_forecaster_
model

TransformedTargetForecaster(steps=[('deseasonalizer',
                                    OptionalPassthrough(passthrough=True,
                                                        transformer=Deseasonalizer())),
                                   ('scaler',
                                    OptionalPassthrough(passthrough=True,
                                                        transformer=TabularToSeriesAdaptor(transformer=StandardScaler()))),
                                   ('forecaster',
                                    NaiveForecaster(strategy='drift'))])

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

Reference

• Forecasting Tutorial