& SzymonStolarski [BUG] fix pmdarima interfaces breaking for X containing more indices than forecasting horizon

[fkiraly]

Fixes #3657.

The bug was caused by pmdarima models breaking when the X passed was strictly larger than the indices in the forecasting horizon.

The example code in #3657 has been added as a test (with minor generalization to cover more estimators).

In the future, we should probably also add test scenarios where X is strictly larger than the forecasting horizon.

Depends on:

• [BUG] fix index error in nullable input test #4474, which fixes a bug that was masked by [BUG] pmdarima estimators break when X contains more indices than the forecasting horizon #3657
• [ENH] allow inclusive/exclusive bounds in get_slice #4483 for slicing X

[fkiraly]

strange: this test fails under 3.10 but not under 3.8

from sktime.forecasting.compose import TransformedTargetForecaster
from sktime.utils.estimator_checks import check_estimator

check_estimator(TransformedTargetForecaster, fixtures_to_run="test_predict_time_index_with_X[TransformedTargetForecaster-1-y:2cols-fh=1-range-int-True]", return_exceptions=False)

[fkiraly]

this seems to be related to statsmodels, not sure what is happening there.

Reverting the change to the statsmodels adapter and moving to separate issue for investigation.

[SzymonStolarski]

Hi @fkiraly,

It's been quite some time since this has been created, however I was wondering if it would be possible to go back to this PR and check the issue with the tests in order to fix #3657?

I've been also playing around with the change in the pmdarima adapter proposed in this PR and it looks like the evaluate funtion works fine. One thing, however, that I noticed is the fact that this change loses the possibility of quickly passing X assumptions without PeriodIndex in predict, e.g.:

y = load_airline()
X = pd.DataFrame({
    'exo_1': y*5,
})

forecaster = ARIMA()
forecaster.fit(y=y, X=X) # Here we are passing data the standard way with `PeriodIndex`

# Here we would like to check the predictions passing quickly exo assumptions
forecaster.predict(fh=[1,2,3], X=pd.Series([123, 213, 321]) 

Will throw an error KeyError: "None of [PeriodIndex(['1961-01', '1961-02', '1961-03'], dtype='period[M]')] are in the [index]".

I assume predict would need to check if data passed in X contains PeriodIndex and based on that decide if:

# PeriodIndex case
fh_abs = fh.to_absolute(self.cutoff).to_pandas()
if X is not None:
    X = X.loc[fh_abs]

or

# No PeriodIndex case
fh_idx = fh.to_indexer(self.cutoff)
if X is not None:
    X = X.loc[fh_idx]

[fkiraly]

I assume predict would need to check

That might be a fix? PR with attempted fixes are appreciated - even if they fail, we learn sth.
Apologies, @SzymonStolarski, for not working on this - the "obvious" fix didn't work and failures were python version dependent.

Maybe we should have a look how this behaves after the pandas 2 compatibility upgrade.

Also FYI @yarnabrina, who has recently been working on the _StatsModelsAdapter, perhaps he can spot sth. Return objects are highly idiosyncratic and not subject to a unified interface in statsmodels (or pmdarima which depends heavily on statsmodels).