Nullable Int64 column changes type after some (cumsum) operations

[has2k1]

Code

import pandas as pd

s1 = pd.Series([1, 2, 3, 4], dtype='Int64')
s2 = pd.Series([1, 2, float('nan'), 3, 4], dtype='Int64')

cs1 = s1.cumsum()
cs2 = s1.cumsum()

print(str(s1.dtype), str(cs1.dtype))
print(str(s2.dtype), str(cs2.dtype))

Output

Int64 object
Int64 object

Expected Output

Int64 Int64
Int64 Int64

Problem description

After an operation like a cumulative sum on a column/series with an Int64 dtype, the dtype of the result is downcast to object. The contents (integers and nans) of the result still qualify for it to have an Int64 dtype.

cummax, cummin and cumprod have the same behaviour.

Output of pd.show_versions()

INSTALLED VERSIONS

commit : None
python : 3.7.4.final.0
python-bits : 64
OS : Linux
OS-release : 5.0.13-arch1-1-ARCH
machine : x86_64
processor :
byteorder : little
LC_ALL : en_US.utf8
LANG : en_US.UTF-8
LOCALE : en_US.UTF-8

pandas : 0.25.1
numpy : 1.17.0
pytz : 2019.2
dateutil : 2.8.0
pip : 19.2.3
setuptools : 41.1.0
Cython : 0.29.13
pytest : 5.1.0
hypothesis : None
sphinx : 2.2.0
blosc : None
feather : None
xlsxwriter : None
lxml.etree : 4.4.1
html5lib : 1.0.1
pymysql : None
psycopg2 : None
jinja2 : 2.10.1
IPython : 7.7.0
pandas_datareader: None
bs4 : 4.8.0
bottleneck : None
fastparquet : None
gcsfs : None
lxml.etree : 4.4.1
matplotlib : 3.1.1
numexpr : 2.7.0
odfpy : None
openpyxl : None
pandas_gbq : None
pyarrow : None
pytables : None
s3fs : None
scipy : 1.3.1
sqlalchemy : None
tables : None
xarray : 0.12.3
xlrd : 1.2.0
xlwt : None
xlsxwriter : None
import pandas as pd

s = pd.Series([1, 2, 3, 4], dtype='Int64')
cs = s.cumsum()

print(str(s.dtype))
print(str(cs.dtype))
import pandas as pd
​

[TomAugspurger]

I don't think that cumulative operations are currently part of the extension array interface.

Likely, we need something similar to ExtensionArray._reduce, but for cumulative methods.

[datajanko]

Python's itertools contain an accumulate method, so I'd suggest ExtensionArray._accumulate and would like to take a shot at this.

[datajanko]

@TomAugspurger would this direction/naming be okay?

Is the assumption correct, that I'd have to e.g. create a nancumsum function in core.nanops.py?

[jorisvandenbossche]

We might need to think a bit more in general about this problem. Adding _accumulate is fine for accumulative methods, but do we want to keep adding such methods for other kinds of functions? (eg the np.round that came up before)
Do we want to eg use the __numpy_function__ protocol for this? Or a similar mechanism specifically for pandas?

Now for this specific case, the numpy protocol is actually not sufficient, as numpy only has cumsum and cumprod, and not cummin/cummax

[mwaskom]

This may be obvious to those who know the implementation details, but it happens with .diff() too (on pandas 0.25.2):

pd.Series([1, 2, 3], dtype="Int16").diff()

0    NaN
1      1
2      1
dtype: object

[jorisvandenbossche]

Series.diff is actually different, and has a custom implementation in pandas, that could be expanded to handle our nullable dtypes. I think it is worth opening a separate issue for that.