I'm doing research related to structural change in stock price time series. In order to test whether some chosen event is a structural break or the series has a structural change, I conduct a few different tests which are

  • Chow Test
  • CUSUM Test
  • SupF Test
  • Bai Perron Test

As the input model, the $AR(3)$ process is used (i.e. $price_{t} = \beta_1 \cdot price_{t-1} + \beta_2 \cdot price_{t-2} + \beta_3 \cdot price_{t-3} + \epsilon_t$).
My concern is that the financial data is highly chaotic so the needed assumptions for the tests are violated.

I wonder what the assumptions that make the structural breaks testing results valid are?

  • $\begingroup$ I think each test may have a (perhaps slightly) different set of assumptions. Have you tried looking at the papers that propose the tests? They would normally contain the assumptions explicitly. $\endgroup$ May 15, 2021 at 13:58

1 Answer 1


In short, you need assumptions that assure that a standard central limit theorem holds: $\sqrt{n} (\hat \beta - \beta) \rightarrow \mathcal{N}(0, V)$. This permits weak correlations and heteroscedasticity provided that a consistent estimate of the covariance matrix $\hat V$ is used. However stronger dependencies (e.g. GARCH) are excluded as are nonstationary or trending regressors which lead to different limiting distributions under the null hypothesis.

Stationary AR(3) models are fine but the residuals should not have remaining autocorrelation in that case.

For details see the corresponding papers. However note that many papers employ somewhat different sets of assumptions and not necessary the broadest assumptions that are possible.


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.