ValueError on make_future_dataframe

[matsuobasho]

What happened + What you expected to happen

Trying to run fcst.make_future_dataframe and getting valuerror.
ValueError: starts must be integers or timestamps, got '<class 'pandas._libs.tslibs.timestamps.Timestamp'>'.

Versions / Dependencies

pandas 2.2.2
mlforecast 0.12.1

Reproduction script

fcst = MLForecast(
    models=lgb.LGBMRegressor(n_jobs=1, random_state=0, verbosity=-1),
    freq='15min',
    lags=[1,2,3,4,5,6,7,8],   # last 2 hours
    lag_transforms={
       1: [ExpandingStd()],
       360: [RollingMean(window_size=14)],
    },
    date_features=['hour', 'dayofweek', 'month'],
)

fcst.fit(train)

fcst.make_future_dataframe(10)

I discovered this while trying to run fcst.predict and getting the same error there.
I can't share the train code, but the ds field is a datetime64[ns, US/Eastern], whereas in the tutorials I see that it's just `datetime64[ns]'.

Could that be the issue?

Btw, I tried to go to the Slack to ask this question (to avoid opening issue), but not able to login with my Google account.

Issue Severity

Medium: It is a significant difficulty but I can work around it.

[jmoralez]

Hey @matsuobasho, thanks for raising this. This should be fixed by Nixtla/utilsforecast#84, which is part of the 0.1.10 release, can you try upgrading to that version (pip install -U utilsforecast)?

[matsuobasho]

Thanks for the response @jmoralez . That worked! Is this a recent change? If not, might want to update which version of utilsforecast is installed, since I had just run pip install mlforecast to install about a month ago.

[jmoralez]

Yes, I just fixed it. I'll bump the utilsforecast version in the next release.