[ARIMA] Incorrect computation of n

[jmoralez]

What happened + What you expected to happen

n here won't have the number of non NA values, but the number minus 1, because the end of the slice isn't included. This was originally reported in #654 since if there's a single sample then n=0 and then we try to take its log.

statsforecast/statsforecast/arima.py

Lines 1212 to 1214 in be9db75

	firstnonmiss = missing_idxs.min()
	lastnonmiss = missing_idxs.max()
	n = np.sum(~missing[firstnonmiss:lastnonmiss])

Versions / Dependencies

1.6.0

Reproduction script

import numpy as np
from statsforecast.models import AutoARIMA

AutoARIMA().fit(np.array([1.0]))

Issue Severity

None

[StatKumar]

@jmoralez Can I work on this issue?

[jmoralez]

Hey @StatKumar. Sure, go ahead! We have a contributing guide here but feel free to ask any questions.

We also have the bug here:

statsforecast/statsforecast/arima.py

Lines 1483 to 1485 in be9db75

	firstnonmiss = np.min(nonmiss_idxs)
	lastnonmiss = np.max(nonmiss_idxs)
	n = np.sum(~missing[firstnonmiss:lastnonmiss])

and here:

statsforecast/statsforecast/arima.py

Lines 1815 to 1817 in be9db75

	firstnonmiss = nonmissing_idxs.min()
	lastnonmiss = nonmissing_idxs.max()
	series_len = int(np.sum(~missing[firstnonmiss:lastnonmiss]))

The fix would be adding a +1 to the lastnonmiss, i.e. lastnonmiss = missing_idxs.max() + 1 so that the last index is considered in the slice below.

[StatKumar]

Thank you @jmoralez