5
$\begingroup$

It's all in the title, I know that we usually use Box-Ljung to test the randomness in a time series (independence of residuals), but I found this post about how to tell if a time series is stationary or not. If this is correct, can someone explain how stationarity is checked using this test?

$\endgroup$

2 Answers 2

8
$\begingroup$

You are quite right that the LB-test should be used for testing lack of serial correlation. In case of a rejection, there is evidence for serial correlation. The rejection however does not allow us to conclude that the process is stationary under the alternative - the LB-test certainly also has power when the series under test is nonstationary.

In that sense, the statement in the linked post

The Ljung-Box test examines whether there is significant evidence for non-zero correlations at lags 1-20. Small p-values (i.e., less than 0.05) suggest that the series is stationary.

is misleading.

This is easy to see from the test statistic $$ LB=T(T+2)\sum_{j=1}^p\frac{\hat\rho_j^2}{T-j} $$ For a random walk as one simple example of a nonstationary process we have that $$Y_T = \sum_{s=1}^{T} \epsilon_s$$ and so $$\gamma_{Tj} = E\left(\sum_{s=1}^T \epsilon_s \sum_{s=1}^{T-j}\epsilon_s\right) = (T-j)\sigma^2$$ Thus, $$\rho_{jT} = \frac{\gamma_{jT}}{\sqrt{\gamma_{0T}}\sqrt{\gamma_{0 (T-j)}}}=\frac{T-j}{\sqrt{T}\sqrt{T-j}}=\frac{\sqrt{T-j}}{\sqrt{T}}= \sqrt{1-\frac{j}{T}}$$ This just recalls the strong persistence properties of a random walk. If the process was "born" long ago ($T\to\infty$), all autocorrelations even tend to one.

Hence, the autocorrelations will also be estimated as $\hat\rho_j\approx1$, $j=1,\ldots,p$, for sufficiently large $T$. Therefore, $LB\to_p\infty$, so that the test will have power equal to 1 in large samples as the test statistic will then virtually always exceed its $\chi^2_p$-critical value. Correspondingly, of course, the $p$-value will be close to zero.

$\endgroup$
1
$\begingroup$

Short answer: no. Lung-Box tests whether the series is white noise. This is a stronger condition than stationarity.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.