Pandas rolling window to return an array
In newer numpy versions there is a sliding_window_view()
.
It provides identical to as_strided()
arrays, but with more transparent syntax.
import pandas as pd
from numpy.lib.stride_tricks import sliding_window_view
x = pd.Series([1, 2, 3, 4, 5, 6, 7, 8, 9])
sliding_window_view(x, 3)
>>>
array([[1, 2, 3],
[2, 3, 4],
[3, 4, 5],
[4, 5, 6],
[5, 6, 7],
[6, 7, 8],
[7, 8, 9]])
But be aware that pandas rolling will add few nans (window_size - 1) at the start because it uses padding. You can check it like this:
x.rolling(3).sum()
>>>
0 NaN
1 NaN
2 6.0
3 9.0
4 12.0
5 15.0
6 18.0
7 21.0
8 24.0
dtype: float64
sliding_window_view(x, 3).sum(axis=1)
>>>
array([ 6, 9, 12, 15, 18, 21, 24])
So real corresponding array should be:
c = np.array([[nan, nan, 1.],
[nan, 1., 2.],
[ 1., 2., 3.],
[ 2., 3., 4.],
[ 3., 4., 5.],
[ 4., 5., 6.],
[ 5., 6., 7.],
[ 6., 7., 8.],
[ 7., 8., 9.]])
c.sum(axis=1)
>>>
array([nan, nan, 6., 9., 12., 15., 18., 21., 24.])
You could use np.stride_tricks
:
import numpy as np
as_strided = np.lib.stride_tricks.as_strided
df
A B
0 -0.272824 -1.606357
1 -0.350643 0.000510
2 0.247222 1.627117
3 -1.601180 0.550903
4 0.803039 -1.231291
5 -0.536713 -0.313384
6 -0.840931 -0.675352
7 -0.930186 -0.189356
8 0.151349 0.522533
9 -0.046146 0.507406
win = 3 # window size
# https://stackoverflow.com/a/47483615/4909087
v = as_strided(df.B, (len(df) - (win - 1), win), (df.B.values.strides * 2))
v
array([[ -1.60635669e+00, 5.10129842e-04, 1.62711678e+00],
[ 5.10129842e-04, 1.62711678e+00, 5.50902812e-01],
[ 1.62711678e+00, 5.50902812e-01, -1.23129111e+00],
[ 5.50902812e-01, -1.23129111e+00, -3.13383794e-01],
[ -1.23129111e+00, -3.13383794e-01, -6.75352179e-01],
[ -3.13383794e-01, -6.75352179e-01, -1.89356194e-01],
[ -6.75352179e-01, -1.89356194e-01, 5.22532550e-01],
[ -1.89356194e-01, 5.22532550e-01, 5.07405549e-01]])
df['C'] = pd.Series(v.tolist(), index=df.index[win - 1:])
df
A B C
0 -0.272824 -1.606357 NaN
1 -0.350643 0.000510 NaN
2 0.247222 1.627117 [-1.606356691642917, 0.0005101298424200881, 1....
3 -1.601180 0.550903 [0.0005101298424200881, 1.6271167809032248, 0....
4 0.803039 -1.231291 [1.6271167809032248, 0.5509028122535129, -1.23...
5 -0.536713 -0.313384 [0.5509028122535129, -1.2312911105674484, -0.3...
6 -0.840931 -0.675352 [-1.2312911105674484, -0.3133837943758246, -0....
7 -0.930186 -0.189356 [-0.3133837943758246, -0.6753521794378446, -0....
8 0.151349 0.522533 [-0.6753521794378446, -0.18935619377656243, 0....
9 -0.046146 0.507406 [-0.18935619377656243, 0.52253255045267, 0.507...
Let's using this pandas approach with a rolling apply trick:
df = pd.DataFrame(np.random.randn(10, 2), columns=list('AB'))
list_of_values = []
df.B.rolling(3).apply(lambda x: list_of_values.append(x.values) or 0, raw=False)
df.loc[2:,'C'] = pd.Series(list_of_values).values
df
Output:
A B C
0 1.610085 0.354823 NaN
1 -0.241446 -0.304952 NaN
2 0.524812 -0.240972 [0.35482336179318674, -0.30495156795594963, -0.24097191924555197]
3 0.767354 0.281625 [-0.30495156795594963, -0.24097191924555197, 0.2816249674055174]
4 -0.349844 -0.533781 [-0.24097191924555197, 0.2816249674055174, -0.5337811449574766]
5 -0.174189 0.133795 [0.2816249674055174, -0.5337811449574766, 0.13379518286397707]
6 2.799437 -0.978349 [-0.5337811449574766, 0.13379518286397707, -0.9783488211443795]
7 0.250129 0.289782 [0.13379518286397707, -0.9783488211443795, 0.2897823417165459]
8 -0.385259 -0.286399 [-0.9783488211443795, 0.2897823417165459, -0.28639931887491943]
9 -0.755363 -1.010891 [0.2897823417165459, -0.28639931887491943, -1.0108913605575793]
Since pandas 1.1
rolling objects are iterable.
For a list of lists:
df['C'] = [window.to_list() for window in df.B.rolling(window=3)]
For a Series of Series's do:
df['C'] = pd.Series(df.B.rolling(window=3))
Also checkout the rolling function for parameters.