Making Multioutput Forecasts without Exogenous Data

[RahulDas-dev]

Hi All,

Thanks for the Beautiful Package, it is really great help for me.

My doubt is why X [ Exogenous Data ] is needed when we are doing multioutput forecasting.
I am really stuck with this issue , Example 1 is working with out X_data but Example 2 is not

Example 1

fhzon = np.arange(ouput_seq_len) + 1
forecaster_pipe = TransformedTargetForecaster(
    [
        ("trnsformer", TabularToSeriesAdaptor(StandardScaler())),
        ("forecaster", make_reduction(LinearRegression(),
                                      window_length=input_seq_len,
                                      strategy="multioutput")
        ),
    ]
)
forecaster_pipe.fit(y_train, X=x_train, fh=fhzon)
y_predict = forecaster_pipe.predict(fh=fhzon)

Example 2

trf_y = TransformedTargetForecaster(
            [
                ("scale_y", TabularToSeriesAdaptor(StandardScaler())),
                ("xtransform", ForecastingPipeline([
                            ("scale_X", TabularToSeriesAdaptor(Id())),
                            ("forecaster", make_reduction(estimator=LinearRegression(),
                                                                            window_length=24,
                                                                            strategy="multioutput"
                                                             )
                            ),
                        ])
                )   
            ])
trf_y.fit(y_train, X=x_train, fh=fhzon)

tt=trf_y.predict(fh=fhzon)

Errors are given bellow

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Cell In[20], line 16
      1 trf_y = TransformedTargetForecaster(
      2             [
      3                 ("scale_y", TabularToSeriesAdaptor(StandardScaler())),
   (...)
     12                 )   
     13             ])
     14 trf_y.fit(y_train, X=x_train, fh=fhzon)
---> 16 tt=trf_y.predict(fh=fhzon)

File ~\AppData\Local\Programs\Python\Python311\Lib\site-packages\sktime\forecasting\base\_base.py:407, in BaseForecaster.predict(self, fh, X)
    405 # we call the ordinary _predict if no looping/vectorization needed
    406 if not self._is_vectorized:
--> 407     y_pred = self._predict(fh=fh, X=X_inner)
    408 else:
    409     # otherwise we call the vectorized version of predict
    410     y_pred = self._vectorize("predict", X=X_inner, fh=fh)

File ~\AppData\Local\Programs\Python\Python311\Lib\site-packages\sktime\forecasting\compose\_pipeline.py:1013, in TransformedTargetForecaster._predict(self, fh, X)
    998 def _predict(self, fh=None, X=None):
    999     """Forecast time series at future horizon.
   1000 
   1001     Parameters
   (...)
   1011         Point predictions
   1012     """
-> 1013     y_pred = self.forecaster_.predict(fh=fh, X=X)
   1014     # inverse transform y_pred
   1015     y_pred = self._get_inverse_transform(self.transformers_pre_, y_pred, X)

File ~\AppData\Local\Programs\Python\Python311\Lib\site-packages\sktime\forecasting\base\_base.py:407, in BaseForecaster.predict(self, fh, X)
    405 # we call the ordinary _predict if no looping/vectorization needed
    406 if not self._is_vectorized:
--> 407     y_pred = self._predict(fh=fh, X=X_inner)
    408 else:
    409     # otherwise we call the vectorized version of predict
    410     y_pred = self._vectorize("predict", X=X_inner, fh=fh)

File ~\AppData\Local\Programs\Python\Python311\Lib\site-packages\sktime\forecasting\compose\_pipeline.py:518, in ForecastingPipeline._predict(self, fh, X)
    503 def _predict(self, fh=None, X=None):
    504     """Forecast time series at future horizon.
    505 
    506     Parameters
   (...)
    516         Point predictions
    517     """
--> 518     X = self._transform(X=X)
    519     return self.forecaster_.predict(fh, X)

File ~\AppData\Local\Programs\Python\Python311\Lib\site-packages\sktime\forecasting\compose\_pipeline.py:693, in ForecastingPipeline._transform(self, X, y)
    691 if self._X is not None and not self.get_tag("ignores-exogeneous-X"):
    692     for _, _, transformer in self._iter_transformers():
--> 693         X = transformer.transform(X=X, y=y)
    694 return X

File ~\AppData\Local\Programs\Python\Python311\Lib\site-packages\sktime\transformations\base.py:533, in BaseTransformer.transform(self, X, y)
    530 self.check_is_fitted()
    532 # input check and conversion for X/y
--> 533 X_inner, y_inner, metadata = self._check_X_y(X=X, y=y, return_metadata=True)
    535 if not isinstance(X_inner, VectorizedDF):
    536     Xt = self._transform(X=X_inner, y=y_inner)

File ~\AppData\Local\Programs\Python\Python311\Lib\site-packages\sktime\transformations\base.py:874, in BaseTransformer._check_X_y(self, X, y, return_metadata)
    820 """Check and coerce X/y for fit/transform functions.
    821 
    822 Parameters
   (...)
    871 ValueError if self.get_tag("requires_y")=True but y is None
    872 """
    873 if X is None:
--> 874     raise TypeError("X cannot be None, but found None")
    876 # skip conversion if it is turned off
    877 if self.get_config()["input_conversion"] != "on":

TypeError: X cannot be None, but found None

I would greatly appreciate any advice, suggestions, or alternative approaches from the community , Thanks in advance

[drackham]

I'm not an expert with Sktime, but I believe what is happening in your second example is you are including transformations on X in the forecasting pipeline. ("scale_X", TabularToSeriesAdaptor(Id())), As such, when the predict step occurs it is expecting X data to transform, but there is none provided, hence the error.

If you're trying to transform X in the train step, but not include X in the predict step, I think you'll need to do the transformation on X outside of the forecasting pipeline.

[fkiraly]

yes, @drackham, I think that is it.

Id already is a time series transformer, so likely TabularToSeriesAdaptor causes the issue (it expects an sklearn transformer inside)

Btw, @RahulDas-dev, it would be great if you could add imports, then this code would be easier to check, i.e., anyone can copy-paste it and run it.