下面是一个使用numba的快速实现(注意类型)。注意它确实包含移位的nan。

import numpy as np
import numba as nb

@nb.jit(nb.float64[:](nb.float64[:],nb.int64),
        fastmath=True,nopython=True)
def moving_average( array, window ):    
    ret = np.cumsum(array)
    ret[window:] = ret[window:] - ret[:-window]
    ma = ret[window - 1:] / window
    n = np.empty(window-1); n.fill(np.nan)
    return np.concatenate((n.ravel(), ma.ravel())) 

这个使用Pandas的答案是从上面改编的，因为rolling_mean不再是Pandas的一部分了

# the recommended syntax to import pandas
import pandas as pd
import numpy as np

# prepare some fake data:
# the date-time indices:
t = pd.date_range('1/1/2010', '12/31/2012', freq='D')

# the data:
x = np.arange(0, t.shape[0])

# combine the data & index into a Pandas 'Series' object
D = pd.Series(x, t)

现在，只需要在窗口大小的数据框架上调用滚动函数，在下面的例子中，窗口大小是10天。

d_mva10 = D.rolling(10).mean()

# d_mva is the same size as the original Series
# though obviously the first w values are NaN where w is the window size
d_mva10[:11]

2010-01-01    NaN
2010-01-02    NaN
2010-01-03    NaN
2010-01-04    NaN
2010-01-05    NaN
2010-01-06    NaN
2010-01-07    NaN
2010-01-08    NaN
2010-01-09    NaN
2010-01-10    4.5
2010-01-11    5.5
Freq: D, dtype: float64

如果有人需要一个简单的解决方案，这里有一个

def moving_average(a,n):
    N=len(a)
    return np.array([np.mean(a[i:i+n]) for i in np.arange(0,N-n+1)])

你可以通过在np.arange(0,N-n+1,step)中添加step参数来改变窗口之间的重叠

实际上，我想要一个稍微不同于公认答案的行为。我正在为sklearn管道构建一个移动平均特征提取器，因此我要求移动平均的输出与输入具有相同的维数。我想要的是让移动平均假设级数保持不变，即[1,2,3,4,5]与窗口2的移动平均将得到[1.5,2.5,3.5,4.5,5.0]。

对于列向量(我的用例)我们得到

def moving_average_col(X, n):
  z2 = np.cumsum(np.pad(X, ((n,0),(0,0)), 'constant', constant_values=0), axis=0)
  z1 = np.cumsum(np.pad(X, ((0,n),(0,0)), 'constant', constant_values=X[-1]), axis=0)
  return (z1-z2)[(n-1):-1]/n

对于数组

def moving_average_array(X, n):
  z2 = np.cumsum(np.pad(X, (n,0), 'constant', constant_values=0))
  z1 = np.cumsum(np.pad(X, (0,n), 'constant', constant_values=X[-1]))
  return (z1-z2)[(n-1):-1]/n

当然，不必假设填充值为常数，但在大多数情况下这样做应该足够了。

这里有许多实现这一点的方法，以及一些基准测试。最好的方法是使用来自其他库的优化代码。瓶颈。Move_mean方法可能是最好的方法。scipy。卷积方法也非常快，可扩展，并且语法和概念简单，但是对于非常大的窗口值不能很好地扩展。numpy。如果你需要一个纯numpy方法，Cumsum方法是很好的。

注意:其中一些(例如:瓶颈。move_mean)不是居中的，并且会转移你的数据。

import numpy as np
import scipy as sci
import scipy.signal as sig
import pandas as pd
import bottleneck as bn
import time as time

def rollavg_direct(a,n): 
    'Direct "for" loop'
    assert n%2==1
    b = a*0.0
    for i in range(len(a)) :
        b[i]=a[max(i-n//2,0):min(i+n//2+1,len(a))].mean()
    return b

def rollavg_comprehension(a,n):
    'List comprehension'
    assert n%2==1
    r,N = int(n/2),len(a)
    return np.array([a[max(i-r,0):min(i+r+1,N)].mean() for i in range(N)]) 

def rollavg_convolve(a,n):
    'scipy.convolve'
    assert n%2==1
    return sci.convolve(a,np.ones(n,dtype='float')/n, 'same')[n//2:-n//2+1]  

def rollavg_convolve_edges(a,n):
    'scipy.convolve, edge handling'
    assert n%2==1
    return sci.convolve(a,np.ones(n,dtype='float'), 'same')/sci.convolve(np.ones(len(a)),np.ones(n), 'same')  

def rollavg_cumsum(a,n):
    'numpy.cumsum'
    assert n%2==1
    cumsum_vec = np.cumsum(np.insert(a, 0, 0)) 
    return (cumsum_vec[n:] - cumsum_vec[:-n]) / n

def rollavg_cumsum_edges(a,n):
    'numpy.cumsum, edge handling'
    assert n%2==1
    N = len(a)
    cumsum_vec = np.cumsum(np.insert(np.pad(a,(n-1,n-1),'constant'), 0, 0)) 
    d = np.hstack((np.arange(n//2+1,n),np.ones(N-n)*n,np.arange(n,n//2,-1)))  
    return (cumsum_vec[n+n//2:-n//2+1] - cumsum_vec[n//2:-n-n//2]) / d

def rollavg_roll(a,n):
    'Numpy array rolling'
    assert n%2==1
    N = len(a)
    rolling_idx = np.mod((N-1)*np.arange(n)[:,None] + np.arange(N), N)
    return a[rolling_idx].mean(axis=0)[n-1:] 

def rollavg_roll_edges(a,n):
    # see https://stackoverflow.com/questions/42101082/fast-numpy-roll
    'Numpy array rolling, edge handling'
    assert n%2==1
    a = np.pad(a,(0,n-1-n//2), 'constant')*np.ones(n)[:,None]
    m = a.shape[1]
    idx = np.mod((m-1)*np.arange(n)[:,None] + np.arange(m), m) # Rolling index
    out = a[np.arange(-n//2,n//2)[:,None], idx]
    d = np.hstack((np.arange(1,n),np.ones(m-2*n+1+n//2)*n,np.arange(n,n//2,-1)))
    return (out.sum(axis=0)/d)[n//2:]

def rollavg_pandas(a,n):
    'Pandas rolling average'
    return pd.DataFrame(a).rolling(n, center=True, min_periods=1).mean().to_numpy()

def rollavg_bottlneck(a,n):
    'bottleneck.move_mean'
    return bn.move_mean(a, window=n, min_count=1)

N = 10**6
a = np.random.rand(N)
functions = [rollavg_direct, rollavg_comprehension, rollavg_convolve, 
        rollavg_convolve_edges, rollavg_cumsum, rollavg_cumsum_edges, 
        rollavg_pandas, rollavg_bottlneck, rollavg_roll, rollavg_roll_edges]

print('Small window (n=3)')
%load_ext memory_profiler
for f in functions : 
    print('\n'+f.__doc__+ ' : ')
    %timeit b=f(a,3)

print('\nLarge window (n=1001)')
for f in functions[0:-2] : 
    print('\n'+f.__doc__+ ' : ')
    %timeit b=f(a,1001)

print('\nMemory\n')
print('Small window (n=3)')
N = 10**7
a = np.random.rand(N)
%load_ext memory_profiler
for f in functions[2:] : 
    print('\n'+f.__doc__+ ' : ')
    %memit b=f(a,3)

print('\nLarge window (n=1001)')
for f in functions[2:-2] : 
    print('\n'+f.__doc__+ ' : ')
    %memit b=f(a,1001)

定时，小窗口(n=3)

Direct "for" loop : 

4.14 s ± 23.7 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

List comprehension : 
3.96 s ± 27.9 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

scipy.convolve : 
1.07 ms ± 26.7 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

scipy.convolve, edge handling : 
4.68 ms ± 9.69 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

numpy.cumsum : 
5.31 ms ± 5.11 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

numpy.cumsum, edge handling : 
8.52 ms ± 11.1 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

Pandas rolling average : 
9.85 ms ± 9.63 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

bottleneck.move_mean : 
1.3 ms ± 12.2 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

Numpy array rolling : 
31.3 ms ± 91.9 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

Numpy array rolling, edge handling : 
61.1 ms ± 55.9 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

定时，大窗口(n=1001)

Direct "for" loop : 
4.67 s ± 34 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

List comprehension : 
4.46 s ± 14.6 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

scipy.convolve : 
103 ms ± 165 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

scipy.convolve, edge handling : 
272 ms ± 1.23 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

numpy.cumsum : 
5.19 ms ± 12.4 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

numpy.cumsum, edge handling : 
8.7 ms ± 11.5 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

Pandas rolling average : 
9.67 ms ± 199 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

bottleneck.move_mean : 
1.31 ms ± 15.7 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

内存，小窗口(n=3)

The memory_profiler extension is already loaded. To reload it, use:
  %reload_ext memory_profiler

scipy.convolve : 
peak memory: 362.66 MiB, increment: 73.61 MiB

scipy.convolve, edge handling : 
peak memory: 510.24 MiB, increment: 221.19 MiB

numpy.cumsum : 
peak memory: 441.81 MiB, increment: 152.76 MiB

numpy.cumsum, edge handling : 
peak memory: 518.14 MiB, increment: 228.84 MiB

Pandas rolling average : 
peak memory: 449.34 MiB, increment: 160.02 MiB

bottleneck.move_mean : 
peak memory: 374.17 MiB, increment: 75.54 MiB

Numpy array rolling : 
peak memory: 661.29 MiB, increment: 362.65 MiB

Numpy array rolling, edge handling : 
peak memory: 1111.25 MiB, increment: 812.61 MiB

内存，大窗口(n=1001)

scipy.convolve : 
peak memory: 370.62 MiB, increment: 71.83 MiB

scipy.convolve, edge handling : 
peak memory: 521.98 MiB, increment: 223.18 MiB

numpy.cumsum : 
peak memory: 451.32 MiB, increment: 152.52 MiB

numpy.cumsum, edge handling : 
peak memory: 527.51 MiB, increment: 228.71 MiB

Pandas rolling average : 
peak memory: 451.25 MiB, increment: 152.50 MiB

bottleneck.move_mean : 
peak memory: 374.64 MiB, increment: 75.85 MiB

通过比较下面的解决方案与使用cumsum of numpy的解决方案，这个解决方案几乎花费了一半的时间。这是因为它不需要遍历整个数组来做cumsum，然后做所有的减法。此外，如果数组很大且数量很大(可能溢出)，cumsum可能是“危险的”。当然，这里也存在危险，但至少我们只把重要的数字加在一起。

def moving_average(array_numbers, n):
    if n > len(array_numbers):
      return []
    temp_sum = sum(array_numbers[:n])
    averages = [temp_sum / float(n)]
    for first_index, item in enumerate(array_numbers[n:]):
        temp_sum += item - array_numbers[first_index]
        averages.append(temp_sum / float(n))
    return averages