NumPy version of "Exponential weighted moving average", equivalent to pandas.ewm().mean()
I think I have finally cracked it!
Here's a vectorized version of numpy_ewma function that's claimed to be producing the correct results from @RaduS's post -
def numpy_ewma_vectorized(data, window):
alpha = 2 /(window + 1.0)
alpha_rev = 1-alpha
scale = 1/alpha_rev
n = data.shape[0]
r = np.arange(n)
scale_arr = scale**r
offset = data[0]*alpha_rev**(r+1)
pw0 = alpha*alpha_rev**(n-1)
mult = data*pw0*scale_arr
cumsums = mult.cumsum()
out = offset + cumsums*scale_arr[::-1]
return out
Further boost
We can boost it further with some code re-use, like so -
def numpy_ewma_vectorized_v2(data, window):
alpha = 2 /(window + 1.0)
alpha_rev = 1-alpha
n = data.shape[0]
pows = alpha_rev**(np.arange(n+1))
scale_arr = 1/pows[:-1]
offset = data[0]*pows[1:]
pw0 = alpha*alpha_rev**(n-1)
mult = data*pw0*scale_arr
cumsums = mult.cumsum()
out = offset + cumsums*scale_arr[::-1]
return out
Runtime test
Let's time these two against the same loopy function for a big dataset.
In [97]: data = np.random.randint(2,9,(5000))
...: window = 20
...:
In [98]: np.allclose(numpy_ewma(data, window), numpy_ewma_vectorized(data, window))
Out[98]: True
In [99]: np.allclose(numpy_ewma(data, window), numpy_ewma_vectorized_v2(data, window))
Out[99]: True
In [100]: %timeit numpy_ewma(data, window)
100 loops, best of 3: 6.03 ms per loop
In [101]: %timeit numpy_ewma_vectorized(data, window)
1000 loops, best of 3: 665 µs per loop
In [102]: %timeit numpy_ewma_vectorized_v2(data, window)
1000 loops, best of 3: 357 µs per loop
In [103]: 6030/357.0
Out[103]: 16.89075630252101
There is around a 17 times speedup!
Updated 08/06/2019
PURE NUMPY, FAST & VECTORIZED SOLUTION FOR LARGE INPUTS
out parameter for in-place computation,
dtype parameter,
index order parameter
This function is equivalent to pandas' ewm(adjust=False).mean(), but much faster. ewm(adjust=True).mean() (the default for pandas) can produce different values at the start of the result. I am working to add the adjust functionality to this solution.
@Divakar's answer leads to floating point precision problems when the input is too large. This is because (1-alpha)**(n+1) -> 0 when n -> inf and alpha -> 1, leading to divide-by-zero's and NaN values popping up in the calculation.
Here is my fastest solution with no precision problems, nearly fully vectorized. It's gotten a little complicated but the performance is great, especially for really huge inputs. Without using in-place calculations (which is possible using the out parameter, saving memory allocation time): 3.62 seconds for 100M element input vector, 3.2ms for a 100K element input vector, and 293µs for a 5000 element input vector on a pretty old PC (results will vary with different alpha/row_size values).
# tested with python3 & numpy 1.15.2
import numpy as np
def ewma_vectorized_safe(data, alpha, row_size=None, dtype=None, order='C', out=None):
"""
Reshapes data before calculating EWMA, then iterates once over the rows
to calculate the offset without precision issues
:param data: Input data, will be flattened.
:param alpha: scalar float in range (0,1)
The alpha parameter for the moving average.
:param row_size: int, optional
The row size to use in the computation. High row sizes need higher precision,
low values will impact performance. The optimal value depends on the
platform and the alpha being used. Higher alpha values require lower
row size. Default depends on dtype.
:param dtype: optional
Data type used for calculations. Defaults to float64 unless
data.dtype is float32, then it will use float32.
:param order: {'C', 'F', 'A'}, optional
Order to use when flattening the data. Defaults to 'C'.
:param out: ndarray, or None, optional
A location into which the result is stored. If provided, it must have
the same shape as the desired output. If not provided or `None`,
a freshly-allocated array is returned.
:return: The flattened result.
"""
data = np.array(data, copy=False)
if dtype is None:
if data.dtype == np.float32:
dtype = np.float32
else:
dtype = np.float
else:
dtype = np.dtype(dtype)
row_size = int(row_size) if row_size is not None
else get_max_row_size(alpha, dtype)
if data.size <= row_size:
# The normal function can handle this input, use that
return ewma_vectorized(data, alpha, dtype=dtype, order=order, out=out)
if data.ndim > 1:
# flatten input
data = np.reshape(data, -1, order=order)
if out is None:
out = np.empty_like(data, dtype=dtype)
else:
assert out.shape == data.shape
assert out.dtype == dtype
row_n = int(data.size // row_size) # the number of rows to use
trailing_n = int(data.size % row_size) # the amount of data leftover
first_offset = data[0]
if trailing_n > 0:
# set temporary results to slice view of out parameter
out_main_view = np.reshape(out[:-trailing_n], (row_n, row_size))
data_main_view = np.reshape(data[:-trailing_n], (row_n, row_size))
else:
out_main_view = out
data_main_view = data
# get all the scaled cumulative sums with 0 offset
ewma_vectorized_2d(data_main_view, alpha, axis=1, offset=0, dtype=dtype,
order='C', out=out_main_view)
scaling_factors = (1 - alpha) ** np.arange(1, row_size + 1)
last_scaling_factor = scaling_factors[-1]
# create offset array
offsets = np.empty(out_main_view.shape[0], dtype=dtype)
offsets[0] = first_offset
# iteratively calculate offset for each row
for i in range(1, out_main_view.shape[0]):
offsets[i] = offsets[i - 1] * last_scaling_factor + out_main_view[i - 1, -1]
# add the offsets to the result
out_main_view += offsets[:, np.newaxis] * scaling_factors[np.newaxis, :]
if trailing_n > 0:
# process trailing data in the 2nd slice of the out parameter
ewma_vectorized(data[-trailing_n:], alpha, offset=out_main_view[-1, -1],
dtype=dtype, order='C', out=out[-trailing_n:])
return out
def get_max_row_size(alpha, dtype=float):
assert 0. <= alpha < 1.
# This will return the maximum row size possible on
# your platform for the given dtype. I can find no impact on accuracy
# at this value on my machine.
# Might not be the optimal value for speed, which is hard to predict
# due to numpy's optimizations
# Use np.finfo(dtype).eps if you are worried about accuracy
# and want to be extra safe.
epsilon = np.finfo(dtype).tiny
# If this produces an OverflowError, make epsilon larger
return int(np.log(epsilon)/np.log(1-alpha)) + 1
The 1D ewma function:
def ewma_vectorized(data, alpha, offset=None, dtype=None, order='C', out=None):
"""
Calculates the exponential moving average over a vector.
Will fail for large inputs.
:param data: Input data
:param alpha: scalar float in range (0,1)
The alpha parameter for the moving average.
:param offset: optional
The offset for the moving average, scalar. Defaults to data[0].
:param dtype: optional
Data type used for calculations. Defaults to float64 unless
data.dtype is float32, then it will use float32.
:param order: {'C', 'F', 'A'}, optional
Order to use when flattening the data. Defaults to 'C'.
:param out: ndarray, or None, optional
A location into which the result is stored. If provided, it must have
the same shape as the input. If not provided or `None`,
a freshly-allocated array is returned.
"""
data = np.array(data, copy=False)
if dtype is None:
if data.dtype == np.float32:
dtype = np.float32
else:
dtype = np.float64
else:
dtype = np.dtype(dtype)
if data.ndim > 1:
# flatten input
data = data.reshape(-1, order)
if out is None:
out = np.empty_like(data, dtype=dtype)
else:
assert out.shape == data.shape
assert out.dtype == dtype
if data.size < 1:
# empty input, return empty array
return out
if offset is None:
offset = data[0]
alpha = np.array(alpha, copy=False).astype(dtype, copy=False)
# scaling_factors -> 0 as len(data) gets large
# this leads to divide-by-zeros below
scaling_factors = np.power(1. - alpha, np.arange(data.size + 1, dtype=dtype),
dtype=dtype)
# create cumulative sum array
np.multiply(data, (alpha * scaling_factors[-2]) / scaling_factors[:-1],
dtype=dtype, out=out)
np.cumsum(out, dtype=dtype, out=out)
# cumsums / scaling
out /= scaling_factors[-2::-1]
if offset != 0:
offset = np.array(offset, copy=False).astype(dtype, copy=False)
# add offsets
out += offset * scaling_factors[1:]
return out
The 2D ewma function:
def ewma_vectorized_2d(data, alpha, axis=None, offset=None, dtype=None, order='C', out=None):
"""
Calculates the exponential moving average over a given axis.
:param data: Input data, must be 1D or 2D array.
:param alpha: scalar float in range (0,1)
The alpha parameter for the moving average.
:param axis: The axis to apply the moving average on.
If axis==None, the data is flattened.
:param offset: optional
The offset for the moving average. Must be scalar or a
vector with one element for each row of data. If set to None,
defaults to the first value of each row.
:param dtype: optional
Data type used for calculations. Defaults to float64 unless
data.dtype is float32, then it will use float32.
:param order: {'C', 'F', 'A'}, optional
Order to use when flattening the data. Ignored if axis is not None.
:param out: ndarray, or None, optional
A location into which the result is stored. If provided, it must have
the same shape as the desired output. If not provided or `None`,
a freshly-allocated array is returned.
"""
data = np.array(data, copy=False)
assert data.ndim <= 2
if dtype is None:
if data.dtype == np.float32:
dtype = np.float32
else:
dtype = np.float64
else:
dtype = np.dtype(dtype)
if out is None:
out = np.empty_like(data, dtype=dtype)
else:
assert out.shape == data.shape
assert out.dtype == dtype
if data.size < 1:
# empty input, return empty array
return out
if axis is None or data.ndim < 2:
# use 1D version
if isinstance(offset, np.ndarray):
offset = offset[0]
return ewma_vectorized(data, alpha, offset, dtype=dtype, order=order,
out=out)
assert -data.ndim <= axis < data.ndim
# create reshaped data views
out_view = out
if axis < 0:
axis = data.ndim - int(axis)
if axis == 0:
# transpose data views so columns are treated as rows
data = data.T
out_view = out_view.T
if offset is None:
# use the first element of each row as the offset
offset = np.copy(data[:, 0])
elif np.size(offset) == 1:
offset = np.reshape(offset, (1,))
alpha = np.array(alpha, copy=False).astype(dtype, copy=False)
# calculate the moving average
row_size = data.shape[1]
row_n = data.shape[0]
scaling_factors = np.power(1. - alpha, np.arange(row_size + 1, dtype=dtype),
dtype=dtype)
# create a scaled cumulative sum array
np.multiply(
data,
np.multiply(alpha * scaling_factors[-2], np.ones((row_n, 1), dtype=dtype),
dtype=dtype)
/ scaling_factors[np.newaxis, :-1],
dtype=dtype, out=out_view
)
np.cumsum(out_view, axis=1, dtype=dtype, out=out_view)
out_view /= scaling_factors[np.newaxis, -2::-1]
if not (np.size(offset) == 1 and offset == 0):
offset = offset.astype(dtype, copy=False)
# add the offsets to the scaled cumulative sums
out_view += offset[:, np.newaxis] * scaling_factors[np.newaxis, 1:]
return out
usage:
data_n = 100000000
data = ((0.5*np.random.randn(data_n)+0.5) % 1) * 100
span = 5000 # span >= 1
alpha = 2/(span+1) # for pandas` span parameter
# com = 1000 # com >= 0
# alpha = 1/(1+com) # for pandas` center-of-mass parameter
# halflife = 100 # halflife > 0
# alpha = 1 - np.exp(np.log(0.5)/halflife) # for pandas` half-life parameter
result = ewma_vectorized_safe(data, alpha)
Just a tip
It is easy to calculate a 'window size' (technically exponential averages have infinite 'windows') for a given alpha, dependent on the contribution of the data in that window to the average. This is useful for example to chose how much of the start of the result to treat as unreliable due to border effects.
def window_size(alpha, sum_proportion):
# Increases with increased sum_proportion and decreased alpha
# solve (1-alpha)**window_size = (1-sum_proportion) for window_size
return int(np.log(1-sum_proportion) / np.log(1-alpha))
alpha = 0.02
sum_proportion = .99 # window covers 99% of contribution to the moving average
window = window_size(alpha, sum_proportion) # = 227
sum_proportion = .75 # window covers 75% of contribution to the moving average
window = window_size(alpha, sum_proportion) # = 68
The alpha = 2 / (window_size + 1.0) relation used in this thread (the 'span' option from pandas) is a very rough approximation of the inverse of the above function (with sum_proportion~=0.87). alpha = 1 - np.exp(np.log(1-sum_proportion)/window_size) is more accurate (the 'half-life' option from pandas equals this formula with sum_proportion=0.5).
In the following example, data represents a continuous noisy signal. cutoff_idx is the first position in result where at least 99% of the value is dependent on separate values in data (i.e. less than 1% depends on data[0]). The data up to cutoff_idx is excluded from the final results because it is too dependent on the first value in data, therefore possibly skewing the average.
result = ewma_vectorized_safe(data, alpha, chunk_size)
sum_proportion = .99
cutoff_idx = window_size(alpha, sum_proportion)
result = result[cutoff_idx:]
To illustrate the problem the above solve you can run this a few times, notice the often-appearing false start of the red line, which is skipped after cutoff_idx:
data_n = 100000
data = np.random.rand(data_n) * 100
window = 1000
sum_proportion = .99
alpha = 1 - np.exp(np.log(1-sum_proportion)/window)
result = ewma_vectorized_safe(data, alpha)
cutoff_idx = window_size(alpha, sum_proportion)
x = np.arange(start=0, stop=result.size)
import matplotlib.pyplot as plt
plt.plot(x[:cutoff_idx+1], result[:cutoff_idx+1], '-r',
x[cutoff_idx:], result[cutoff_idx:], '-b')
plt.show()
note that cutoff_idx==window because alpha was set with the inverse of the window_size() function, with the same sum_proportion.
This is similar to how pandas applies ewm(span=window, min_periods=window).
Fastest EWMA 23x pandas
The question is strictly asking for a numpy solution, however, it seems that the OP was actually just after a pure numpy solution to speed up runtime.
I solved a similar problem but instead looked towards numba.jit which massively speeds the compute time
In [24]: a = np.random.random(10**7)
...: df = pd.Series(a)
In [25]: %timeit numpy_ewma(a, 10) # /a/42915307/4013571
...: %timeit df.ewm(span=10).mean() # pandas
...: %timeit numpy_ewma_vectorized_v2(a, 10) # best w/o numba: /a/42926270/4013571
...: %timeit _ewma(a, 10) # fastest accurate (below)
...: %timeit _ewma_infinite_hist(a, 10) # fastest overall (below)
4.14 s ± 116 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
991 ms ± 52.2 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
396 ms ± 8.39 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
181 ms ± 1.01 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
39.6 ms ± 979 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
Scaling down to smaller arrays of a = np.random.random(100) (results in the same order)
41.6 µs ± 491 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
945 ms ± 12 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
16 µs ± 93.5 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)
1.66 µs ± 13.7 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)
1.14 µs ± 5.57 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)
It is also worth pointing out that my functions below are identically aligned to the pandas (see the examples in docstr), whereas a few of the answers here take various different approximations. For example,
In [57]: print(pd.DataFrame([1,2,3]).ewm(span=2).mean().values.ravel())
...: print(numpy_ewma_vectorized_v2(np.array([1,2,3]), 2))
...: print(numpy_ewma(np.array([1,2,3]), 2))
[1. 1.75 2.61538462]
[1. 1.66666667 2.55555556]
[1. 1.18181818 1.51239669]
The source code which I have documented for my own library
import numpy as np
from numba import jit
from numba import float64
from numba import int64
@jit((float64[:], int64), nopython=True, nogil=True)
def _ewma(arr_in, window):
r"""Exponentialy weighted moving average specified by a decay ``window``
to provide better adjustments for small windows via:
y[t] = (x[t] + (1-a)*x[t-1] + (1-a)^2*x[t-2] + ... + (1-a)^n*x[t-n]) /
(1 + (1-a) + (1-a)^2 + ... + (1-a)^n).
Parameters
----------
arr_in : np.ndarray, float64
A single dimenisional numpy array
window : int64
The decay window, or 'span'
Returns
-------
np.ndarray
The EWMA vector, same length / shape as ``arr_in``
Examples
--------
>>> import pandas as pd
>>> a = np.arange(5, dtype=float)
>>> exp = pd.DataFrame(a).ewm(span=10, adjust=True).mean()
>>> np.array_equal(_ewma_infinite_hist(a, 10), exp.values.ravel())
True
"""
n = arr_in.shape[0]
ewma = np.empty(n, dtype=float64)
alpha = 2 / float(window + 1)
w = 1
ewma_old = arr_in[0]
ewma[0] = ewma_old
for i in range(1, n):
w += (1-alpha)**i
ewma_old = ewma_old*(1-alpha) + arr_in[i]
ewma[i] = ewma_old / w
return ewma
@jit((float64[:], int64), nopython=True, nogil=True)
def _ewma_infinite_hist(arr_in, window):
r"""Exponentialy weighted moving average specified by a decay ``window``
assuming infinite history via the recursive form:
(2) (i) y[0] = x[0]; and
(ii) y[t] = a*x[t] + (1-a)*y[t-1] for t>0.
This method is less accurate that ``_ewma`` but
much faster:
In [1]: import numpy as np, bars
...: arr = np.random.random(100000)
...: %timeit bars._ewma(arr, 10)
...: %timeit bars._ewma_infinite_hist(arr, 10)
3.74 ms ± 60.2 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
262 µs ± 1.54 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
Parameters
----------
arr_in : np.ndarray, float64
A single dimenisional numpy array
window : int64
The decay window, or 'span'
Returns
-------
np.ndarray
The EWMA vector, same length / shape as ``arr_in``
Examples
--------
>>> import pandas as pd
>>> a = np.arange(5, dtype=float)
>>> exp = pd.DataFrame(a).ewm(span=10, adjust=False).mean()
>>> np.array_equal(_ewma_infinite_hist(a, 10), exp.values.ravel())
True
"""
n = arr_in.shape[0]
ewma = np.empty(n, dtype=float64)
alpha = 2 / float(window + 1)
ewma[0] = arr_in[0]
for i in range(1, n):
ewma[i] = arr_in[i] * alpha + ewma[i-1] * (1 - alpha)
return ewma