bug of autocorrelation plot in matplotlib‘s plt.acorr?

Question

I am plotting autocorrelation with python. I used three ways to do it: 1. pandas, 2. matplotlib, 3. statsmodels. I found the graph I got from matplotlib is not consistent with t

Answer 1

This is a result of different common definitions between statistics and signal processing. Basically, the signal processing definition assumes that you're going to handle the detrending. The statistical definition assumes that subtracting the mean is all the detrending you'll do, and does it for you.

First off, let's demonstrate the problem with a stand-alone example:

import numpy as np
import matplotlib.pyplot as plt

import pandas as pd
from statsmodels.graphics import tsaplots

def label(ax, string):
    ax.annotate(string, (1, 1), xytext=(-8, -8), ha='right', va='top',
                size=14, xycoords='axes fraction', textcoords='offset points')

np.random.seed(1977)
data = np.random.normal(0, 1, 100).cumsum()

fig, axes = plt.subplots(nrows=4, figsize=(8, 12))
fig.tight_layout()

axes[0].plot(data)
label(axes[0], 'Raw Data')

axes[1].acorr(data, maxlags=data.size-1)
label(axes[1], 'Matplotlib Autocorrelation')

tsaplots.plot_acf(data, axes[2])
label(axes[2], 'Statsmodels Autocorrelation')

pd.tools.plotting.autocorrelation_plot(data, ax=axes[3])
label(axes[3], 'Pandas Autocorrelation')

# Remove some of the titles and labels that were automatically added
for ax in axes.flat:
    ax.set(title='', xlabel='')
plt.show()

So, why the heck am I saying that they're all correct? They're clearly different!

Let's write our own autocorrelation function to demonstrate what plt.acorr is doing:

def acorr(x, ax=None):
    if ax is None:
        ax = plt.gca()
    autocorr = np.correlate(x, x, mode='full')
    autocorr /= autocorr.max()

    return ax.stem(autocorr)

If we plot this with our data, we'll get a more-or-less identical result to plt.acorr (I'm leaving out properly labeling the lags, simply because I'm lazy):

fig, ax = plt.subplots()
acorr(data)
plt.show()

This is a perfectly valid autocorrelation. It's all a matter of whether your background is signal processing or statistics.

This is the definition used in signal processing. The assumption is that you're going to handle detrending your data (note the detrend kwarg in plt.acorr). If you want it detrended, you'll explictly ask for it (and probably do something better than just subtracting the mean), and otherwise it shouldn't be assumed.

In statistics, simply subtracting the mean is assumed to be what you wanted to do for detrending.

All of the other functions are subtracting the mean of the data before the correlation, similar to this:

def acorr(x, ax=None):
    if ax is None:
        ax = plt.gca()

    x = x - x.mean()

    autocorr = np.correlate(x, x, mode='full')
    autocorr /= autocorr.max()

    return ax.stem(autocorr)

fig, ax = plt.subplots()
acorr(data)
plt.show()

However, we still have one large difference. This one is purely a plotting convention.

In most signal processing textbooks (that I've seen, anyway), the "full" autocorrelation is displayed, such that zero lag is in the center, and the result is symmetric on each side. R, on the other hand, has the very reasonable convention to display only one side of it. (After all, the other side is completely redundant.) The statistical plotting functions follow the R convetion, and plt.acorr follows what Matlab does, which is the opposite convention.

Basically, you'd want this:

def acorr(x, ax=None):
    if ax is None:
        ax = plt.gca()

    x = x - x.mean()

    autocorr = np.correlate(x, x, mode='full')
    autocorr = autocorr[x.size:]
    autocorr /= autocorr.max()

    return ax.stem(autocorr)

fig, ax = plt.subplots()
acorr(data)
plt.show()