forecasting - get fitted values

[eyalshafran]

When using statsmodel it is possible to access the fitted values after the model is fit.
For example:

from sktime.datasets import load_airline
from statsmodels.tsa.api import ExponentialSmoothing as _ExponentialSmoothing
from sktime.forecasting.model_selection import temporal_train_test_split

y = load_airline()
y_train, y_test = temporal_train_test_split(y, test_size=36)
model = _ExponentialSmoothing(y_train, seasonal_periods=12, trend='add', seasonal='add', use_boxcox=True, initialization_method="estimated").fit()
model.fittedvalues

I can't figure out how to access the fitted values using sktime. Is it possible to get them?

[mloning]

Hi @eyalshafran you have two options:

Using the sktime interface for in-sample predictions (this should work for all forecasters):

from sktime.datasets import load_airline
from sktime.forecasting.exp_smoothing import ExponentialSmoothing
from sktime.forecasting.model_selection import temporal_train_test_split
from sktime.forecasting.base import ForecastingHorizon

y = load_airline()
y_train, y_test = temporal_train_test_split(y, test_size=36)
model = ExponentialSmoothing(sp=12, trend='add', seasonal='add')
model.fit(y_train)
fh = ForecastingHorizon(y_train.index, is_relative=False)  # in-sample forecasting horizon
model.predict(fh)

Accessing the statsmodels used internally (this will only work for forecasters based on statsmodels):

model._fitted_forecaster.fittedvalues

[mloning]

Hi @eyalshafran, yes, possibly. This is an error that comes from the pmdarima package as far as I understand. So perhaps it's worth asking there (there may be an issue on this already, open or closed) and link the discussion here?

[eyalshafran]

@mloning if I run the code in pdmarima I don't get an error message:

from pmdarima.arima import ARIMA as _ARIMA
from sktime.datasets import load_airline
from sktime.forecasting.model_selection import temporal_train_test_split

y = load_airline()
y_train, y_test = temporal_train_test_split(y, test_size=36)
model = _ARIMA(order=(1, 1, 1), seasonal_order=(1, 1, 0, 12))
model.fit(y_train)
model.predict_in_sample()

I believe the error is caused by calling the predict_in_sample with a start value smaller than d. This is how the pdmarima adapter calls the predict_in_sample. It is interesting that if you don't specify the start at all there is no error.

[mloning]

Alright, could you raise an issue about this with a reproducible example as above, once in sktime and once using pdmarima directly. I haven't fully understood yet where the error comes from, but would appreciate a PR if you think you it's related to sktime rather than pmdarima!

[eyalshafran]

I'll raise this issue with pdmarima. The error is a result of how predict_in_sample is called. There is inconsistency between calling it without any variables and calling it with start = 0 (which is what the pdmarima adapter does in sktime). Those two ways should be consistent.

[mloning]

Thanks @eyalshafran - that sounds like we need to adjust the start value that we pass to predict_in_sample depending on the forecasting horizon fh and model hyperparameter d. Would you be interested in opening a PR?