Pandas rolling function with specific numeric span? - pandas

Question

As of Pandas 0.18.0, it is possible to have a variable rolling window size for time-series by specifying a time span. For example, the code for summation over a 2-second window in dataframe dft looks like this:
dft.rolling('2s').sum()
It is possible to do the same with non-datetime spans?
For example, given a dataframe that looks like this:
A B
0 1 1
1 2 2
2 3 3
3 5 5
4 6 6
5 7 7
6 10 10
Is it possible to specify a window span of say 3 on column 'A' and have the sum of column 'B' calculated, so that the output looks something like:
A B
0 1 NaN
1 2 NaN
2 3 5
3 5 10
4 6 14
5 7 18
6 10 17

Answer 1

Not with rolling(). See the documentation for the window argument:
[A variable-sized window] is only valid for datetimelike indexes.
Full text:
window : int, or offset
Size of the moving window. This is the number of observations used for calculating the statistic. Each window will be a fixed size.
If its an offset then this will be the time period of each window. Each window will be a variable sized based on the observations included in the time-period. This is only valid for datetimelike indexes.

Answer 2

Here's a workaround if you're interested.
df = pd.DataFrame({'A' : np.arange(10),
'B' : np.arange(10,20)},
index=[1,2,3,5,8,9,11,14,19,20])
def var_window(df, size, min_periods=None):
"""Operates on the index."""
result = []
df = df.sort_index()
for i in df.index:
start = i - size + 1
res = df.loc[start:i].sum().tolist()
result.append(res)
result = pd.DataFrame(result, index=df.index)
if min_periods:
result.loc[:min_periods - 1] = np.nan
return result
print(var_window(df, size=3, min_periods=3, inclusive=True))
0 1
1 NaN NaN
2 NaN NaN
3 3.0 33.0
5 5.0 25.0
8 4.0 14.0
9 9.0 29.0
11 11.0 31.0
14 7.0 17.0
19 8.0 18.0
20 17.0 37.0
Explanation: loop through the index. At each value, truncate the DataFrame to the trailing window size. Here 'size' is not a count, but rather a range as you have defined it.
In the above, at the index value of 8, you're summing the values of A for which the index is 8, 7, or 6. (I.e. > 8 - 3 + 1). The only index value that falls within that range is 8, so the sum is simply the value from the original frame. Comparatively, for the index value of 11, the sum will include values for 9 and 11 (5 + 6 = 11, the resulting sum for A).
Compare this with standard rolling ops:
print(df.rolling(window=3).sum())
A B
1 NaN NaN
2 NaN NaN
3 3.0 33.0
5 6.0 36.0
8 9.0 39.0
9 12.0 42.0
11 15.0 45.0
14 18.0 48.0
19 21.0 51.0
20 24.0 54.0
If I'm misinterpreting your question, let me know how. It's admittedly significantly slower:
%timeit df.rolling(window=3).sum()
1000 loops, best of 3: 627 µs per loop
%timeit var_window(df, size=3, min_periods=3)
100 loops, best of 3: 3.59 ms per loop