I try to test a time series for white noise. The ultimate goal is to show that scaling volatility from daily to longer time periods by the square-root of time rule is justified.

Fore white noise I found the classical tests such as the Ljung-Box test. An example of its application can be found in Forecasting: principles and practice.

What I was wondering is the following: tests like the Ljung-Box test look at auto-correlations which need to be (close to) zero for White noise.

On the other hand we need a constant volatility thus homoscedasticity. Why is it that the "white noise" tests do not test this property of white noise? Or do they test it indirectly?

How can I test homoscedasticity in the time series setting. Tests such as Breusch-Pagan need covariates if I see it correctly.


1 Answer 1


"How To Test for a homogenous error process" : In the absence of the need for a Box-Cox transformation use the residuals and form k distinct subsets .....say k=5 and N = # of observations = 100 we have 1-20 ; 21-40 ; 41-60 ; 61-80 ; 81-100 . Compute the variance in each subset and form an F test comparing successive groups and conclude about a possible breakpoint. Re-estimate model with weights derived from the max F and test to see if another of the k breakpoints is significant . If so integrate/combine the weights. This ultimately yields a set of N weights leading to Weighted Least Squares or GLS . This is how I programmed AUTOBOX to solve the problem.

  • $\begingroup$ Thanjk you! You mean, I calculate the volatility on each of the subsets? This came to my mind too but isn't this too "easy"? :) $\endgroup$
    – Richi W
    Feb 23, 2018 at 15:13
  • $\begingroup$ yes calculate the variance of each subgroup ;;; $\endgroup$
    – IrishStat
    Feb 23, 2018 at 15:27

Your Answer

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

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