Numpy std function seems to be giving different ddof values in pd.apply depending on whether or not the apply argument is an iterable.
In other words, the line
pd.DataFrame([0, 1]).apply(np.std)
Produces different value from
pd.DataFrame([0, 1]).apply([np.std])
The first one is the correct for numpy with ddof=0 while the second one is the Pandas default with ddof=1.
Why is that the case? What changes from a single argument to a list?
I guess this is because under the covers of pd.DataFrame.apply it uses pd.DataFrame.aggregate if you pass a list-like object, and in aggregate it's replaced with Series.std(), which has the default ddof=1 value. Hence the result:
>>> pd.DataFrame([0, 1]).apply([np.std])
0
std 0.707107
In the other case when you pass np.std as a regular function, then it doesn't go through any replacement, and np.std has ddof=0 as default, so this becomes:
>>> pd.DataFrame([0, 1]).apply(np.std)
0 0.5
dtype: float64
This is actually a "wontfix" issue: see here. If you want to skip the replacement of numpy functions into pandas functions use it as follows (from the linked github issue):
.groupby(...).agg(lambda x: np.std(x))