API/BUG: center=True in expanding operations returns non-sensical results

[jorisvandenbossche]

What is center=True supposed to do in df.expanding()... operations?

Using the example from the docstring of .expanding():

In [103]: df = DataFrame({'B': [0, 1, 2, np.nan, 4]})

In [105]: df.expanding(2).sum()
Out[105]: 
     B
0  NaN
1  1.0
2  3.0
3  3.0
4  7.0

Now adding center=True:

In [106]: df.expanding(2, center=True).sum()
Out[106]: 
     B
0  3.0
1  3.0
2  7.0
3  7.0
4  6.0

Two observations: it seems it just shifted the first rows (which is not the same as centering) and I don't know where the last values are coming from (and they should certainly not decrease again in an expanding sum with only positive values).

Further, I also cannot think of what it actually should do. If we really want to center it, that would mean returning with a new index (as the real center would only increase with 0.5 and be something like [0, 0.5, 1, 1.5, 2] for those data).

So maybe we should rather remove this keyword for expanding ?

[jorisvandenbossche]

Hmm, I see that we in the past removed the center keyword in the pd.expanding_ functions, exactly because it did not make sense (#7934). But it seems with the refactor to the expanding() method, it was introduced again

[jreback]

yes center doesn’t make sense in context of expanding

[pulkitmaloo]

I'd like to work on this

[gfyoung]

Go for it! No need to ask for permission unless someone else is working on it.

[dragosthealex]

Is anyone still working on this?

[jorisvandenbossche]

I don't think so (unless @stillmatic did work on it? If you do, best to say so on this issue), so feel free to take a look at it!

[stillmatic]

i started working on this issue but don't think my solution is correct/finished. Haven't had time to finish (there are many dependencies). If you have a good idea go ahead, otherwise I might be able to find time and get this done

[dragosthealex]

@stillmatic in the commit you removed center parameter from rolling method as well, which I believe introduced most broken dependencies. But I believe in that one, centering makes sense.

@jorisvandenbossche I believe in this example center should do nothing if it's supposed to be consistent with functionality of rolling.
As far as I understand, when center=False, the result of the operation applied on the window is placed at the index of last element in window. e.g. df = DataFrame({'B': [0, 1, 2, 3, 4, 5]})
df.rolling(3).sum() gives [NaN, NaN, 3.0, 6.0, 9.0, 12.0]

When setting center=True, the result is placed at the index of element in center of the window (index in window / 2), effectively shifting the results back.
df.rolling(3, center=True).sum() gives [NaN, 3.0, 6.0, 9.0, 12.0, NaN]

df.rolling(2).sum() gives same result as df.rolling(2, center=True).sum(), because the center of window of len=2, is index=1 -> second element in window, which is also the last. Thus, df.expanding(2).sum()should give same result asdf.expanding(2, center=True).sum()`, so there must be a bug somewhere.
Please let me know if you think I got it wrong.

In terms of whether it's actually useful for expanding I'm not sure as I am not really familiar, but it should have the same effect as in rolling e.g.
df.expanding(3).sum() gives [NaN, NaN, 3.0, 6.0, 10.0, 15.0]
And df.expanding(3, center=True).sum() should give [NaN, 3.0, 6.0, 10.0, 15.0, NaN] right?

[sumanau7]

take

[MBrouns]

It seems like center is indeed a weird option for an expanding window. I just ran the test suite and there are no tests breaking when I remove the center option from expanding (except for one that checks whether there's validation on the center parameter). Also, a code search on GitHub doesn't reveal any instances where center is used in conjuction with expanding.

Combined with the current output simply being wrong, I think that means we can move forward with deprecating center. I'll submit a PR.

[MBrouns]

take