window-ops 0.0.4

Naive implementations of window operations such as rolling and expanding.

Window ops

Naive and fast implementations of common window operations.

This library is intended to be used as a replacement to pd.rolling and pd.expanding to gain a speedup by operating on numpy arrays and avoiding input checks.

Install

pip install window-ops

How to use

Rolling

Simply use rolling_{op} with a one dimensional np.ndarray by specifying the window size and the minimum number of samples to compute the operation (by default min_samples equals window_size). The result will have min_samples - 1 np.nan's at the beggining of the array.

import random

import numpy as np
import pandas as pd
from window_ops.rolling import *
from window_ops.expanding import *
from window_ops.ewm import *

np.random.seed(0)
random.seed(0)
y = np.random.rand(10)
window_size = 3
y

array([0.5488135 , 0.71518937, 0.60276338, 0.54488318, 0.4236548 ,
       0.64589411, 0.43758721, 0.891773  , 0.96366276, 0.38344152])

rolling_mean(y, window_size=window_size)

array([       nan,        nan, 0.62225544, 0.62094533, 0.5237671 ,
       0.53814405, 0.5023787 , 0.6584181 , 0.764341  , 0.7462924 ],
      dtype=float32)

ys = pd.Series(y)
ys.rolling(window_size).mean().values

array([       nan,        nan, 0.62225542, 0.62094531, 0.52376712,
       0.53814403, 0.50237871, 0.65841811, 0.76434099, 0.74629243])

big_y = np.random.rand(1_000)

%timeit rolling_mean(big_y, window_size=8)

2.53 µs ± 84 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

big_ys = pd.Series(big_y)

%timeit big_ys.rolling(8).mean()

305 µs ± 11.9 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

Expanding

Simply use expanding_{op} with a one dimensional np.ndarray. For expanding_std the first value in the output array is np.nan, for all the other operations a full array is returned.

y

array([0.5488135 , 0.71518937, 0.60276338, 0.54488318, 0.4236548 ,
       0.64589411, 0.43758721, 0.891773  , 0.96366276, 0.38344152])

expanding_mean(y)

array([0.5488135 , 0.63200146, 0.62225544, 0.60291237, 0.5670608 ,
       0.5801997 , 0.5598265 , 0.6013198 , 0.64158016, 0.6157663 ],
      dtype=float32)

ys.expanding().mean().values

array([0.5488135 , 0.63200144, 0.62225542, 0.60291236, 0.56706085,
       0.58019972, 0.55982651, 0.60131982, 0.64158015, 0.61576628])

%timeit expanding_mean(big_y)

1.87 µs ± 50.5 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)

%timeit big_ys.expanding().mean()

325 µs ± 13.5 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

EWM

Simply use ewm_{op} with a one dimensional np.ndarray and the smoothing parameter alpha. Currently only ewm_mean is implemented.

y

array([0.5488135 , 0.71518937, 0.60276338, 0.54488318, 0.4236548 ,
       0.64589411, 0.43758721, 0.891773  , 0.96366276, 0.38344152])

alpha = random.random()
alpha

0.8444218515250481

ewm_mean(y, alpha=alpha)

array([0.5488135 , 0.68930495, 0.6162273 , 0.55598277, 0.44424215,
       0.6145215 , 0.46511433, 0.8253942 , 0.9421512 , 0.47036454],
      dtype=float32)

ys.ewm(alpha=alpha, adjust=False).mean().values

array([0.5488135 , 0.68930492, 0.61622735, 0.55598278, 0.44424214,
       0.61452147, 0.46511432, 0.82539423, 0.9421512 , 0.47036454])

%timeit ewm_mean(big_y, alpha)

5.56 µs ± 62.6 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

%timeit big_ys.ewm(alpha=alpha, adjust=False).mean().values

302 µs ± 13 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)