2
$\begingroup$

I am using the Ljung-Box test from the python statsmodel package to test if there are autocorrelations in a time series. To make sure I understand the test, I am performing it on white noise and checking that the null hypothesis H0 that there is no autocorrelation is being rejected a number of times consistent with the chosen p-value.

However, I am finding that, for large lags, H0 is being rejected more than expected on average. I wrote the simple code:

import numpy as np
import statsmodels
import statsmodels.api as sm
from statsmodels import stats

N = 10000
T = 100

x = np.random.normal(0, 1, T)
_, pvalues = statsmodels.stats.diagnostic.acorr_ljungbox(x)
ret = np.zeros(shape=pvalues.shape)

for i in range(N):
    x = np.random.normal(0,1,T)
    _, pvalues = statsmodels.stats.diagnostic.acorr_ljungbox(x)
    ret += (pvalues<0.05).astype(int)

ret /= N
print(ret)

The typical output of the code is like the following:

[ 0.0522 0.0514 0.052 0.0512 0.0539 0.0549 0.0573 0.0592 0.0605 0.0625 0.0638 0.067 0.0674 0.0694 0.0703 0.0719 0.0715 0.0734 0.0749 0.0747 0.0769 0.0765 0.077 0.0774 0.0797 0.079 0.0817 0.083 0.0857 0.0873 0.0874 0.088 0.0891 0.0914 0.0909 0.0914 0.0926 0.0934 0.0936 0.0934]

So the H0 tends to be consistently rejected more often than 5% of the times for larger lags. Am I misinterpreting the meaning of this test or making any other conceptual mistake?

EDIT: Based on the comment below I changed the number of observation and made plots of the fraction of times the null hypothesis is rejected.

plotTypeIerror

Therefore it looks like for a set of observation of size 100, testing beyond a lag of around 5-10 incurs in a larger type I error

$\endgroup$
3
  • $\begingroup$ Try with larger sample size T, and check that it doesn't use too many lags. I don't know how good the small sample behavior of ljung-box is. $\endgroup$ – Josef Apr 10 '18 at 20:39
  • $\begingroup$ Thanks @user333700 . I was a bit fooled by the documentation of the statsmodel acorr_ljungbox function, which claims to set the maximum of tested lags to 12*(nobs/100)^{1/4} , which I thought might be theoretically motivated. However I realized that the default maxlags is always 40 regardless of the size of the input. Here is the docs statsmodels.org/dev/generated/… $\endgroup$ – chubecca Apr 11 '18 at 4:25
  • $\begingroup$ see also stats.stackexchange.com/questions/200267/… for some comments and recommendation to choose the number of lags. $\endgroup$ – Josef Apr 11 '18 at 12:07

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.