API/BUG: Series.argmin/max with all-NaN data returns -1 ?

[jorisvandenbossche]

Ran into this while looking at #27801.

Both numpy and pandas raise an error for empty arrays/Series in argmin and argmax:

>>> pd.Series([], dtype=float).argmin()   
...
ValueError: attempt to get argmin of an empty sequence

>>> np.array([], dtype=float).argmin()  
...
ValueError: attempt to get argmin of an empty sequence

However, when having all NaN data, we see a different behaviour:

>>> pd.Series([np.nan, np.nan], dtype=float).argmin() 
-1

>>> np.array([np.nan, np.nan], dtype=float).argmin() 
0

Does somebody have an explanation of why this would return -1 ?

In principle, in pandas, argmin has a skipna=True keyword, so for pandas I would expect that an all-NaN Series behaves the same as an empty Series.

[jorisvandenbossche]

So this is quite explicit in the code of nanops.nanargmin/max:

pandas/pandas/core/nanops.py

Lines 1171 to 1176 in 07402f6

	if skipna:
	if mask.all():
	result = -1
	else:
	if mask.any():
	result = -1

and those lines haven't been touched in 8 years .. c1307b6

This behaviour is used in idxmin/max:

pandas/pandas/core/series.py

Lines 2005 to 2008 in 07402f6

	i = nanops.nanargmin(self._values, skipna=skipna)
	if i == -1:
	return np.nan
	return self.index[i]

The logic behind "-1" might be that in general in pandas indexers, -1 is used for missing values (and argmin kind of returns an indexer). But still, that doesn't make much sense to me for a user-facing argmin/max.

[jreback]

if you recall we separated out argmax from idxmax a while back

since argmax is not that useful in pandas (you almost always want idxmax) it’s likely understated

not averse to changing to more common behavior to raise for argmax

however min/max imho should just work in empty (no raise) as that is a common pandas pattern

[jorisvandenbossche]

if you recall we separated out argmax from idxmax a while back

Yes, I know. But, which actually means that the new user-facing behaviour of argmin/argmax is rather recent (only from pandas 1.0), which is maybe a good reason that we can still fix this as a bug instead of deprecating.

Now, the actual behaviour of nanops.nanargmin/max is from long before this idx/arg split. It even stems from before pandas 0.13, when Series stopped being an ndarray subclass and changed the behaviour of argmin/max. Now, I suppose before 1.0, the nanops version was only used internally (eg in the idxmin/max implementation, as shown above)

[jorisvandenbossche]

Now, I suppose before 1.0, the nanops version was only used internally (eg in the idxmin/max implementation, as shown above)

On second thought, that's not fully true, since the Index classes already have been using the nanops version in their argmin/max methods (and does already returns -1 for a long time in this case).