My question comes from this paper (p. 10), where the authors say:

The $\ln(\hat{p_t}) - \ln(p_t) \sim AR(1)$ condition expresses that the LPPLS fitting residuals can be modeled by a mean-reversal Ornstein-Uhlenbeck (O-U) process when the logarithmic price in the bubble regime is attributed to a deterministic LPPLS component. Since the test for the O-U property of LPPLS fitting residuals can be translated into an AR(1) test for the corresponding residuals, both the Phillips-Perron unit-root test and the Dickey-Fuller unit-root test are used to check the O-U property of LPPLS fitting residuals. In this study, the 5% significant level is applied in the tests.

$\ln(\hat{p_t}) - \ln(p_t)$ are the residuals between the log-values of a fitted curve ($\hat{p}_t$) and the observed log-values ($p_t$).

From my understanding the authors use the Phillips-Perron (PP) and Dickey-Fuller (DF) unit-root tests to test if the residuals $\ln(\hat{p_t}) - \ln(p_t)$ satisfy an AR(1) or O-U process.

I know that the unit-root test checks, if for a given AR(1) process:

$$ y_t = \theta y_{t-1} + \epsilon_t $$ one of the following conditions hold:

$$ \begin{align} H_0: \theta=1 \\ H_A: \theta<1 \end{align} $$ If $H_0$ is accepted, the process is non-stationary.

I'm also aware, that an AR(1) process is the time-discrete version of an Ornstein-Uhlenbeck process.

The actual question: I can't see how the two tests (PP and DF) can tell me if the residuals $\ln(\hat{p_t}) - \ln(p_t)$ satisfy an AR(1) or O-U process, since from my understanding they just check if a specific property (non-stationarity) of the processes is satisfied.

In other words: For what purpose are the authors using the PP and DF tests? It seems to me they are using both tests to test wether the residuals can be modeled with a AR(1) / O-U process. But from my understanding that is not what can be tested using the PP / DK tests.

  • $\begingroup$ Quick answer: Yes, you are correct these test are unit root tests. $\endgroup$ Commented Jun 29, 2019 at 18:54
  • $\begingroup$ @HartoSaarinen I know that they are unit root tests, but why should they tell me if the data is actually an AR(1) or O-U process? $\endgroup$
    – WolfgangP
    Commented Jun 29, 2019 at 19:03
  • $\begingroup$ They don’t tell you that. They assume the AR model. Of course you could use a different model and then you have some different tests. But AR (and/or MA ) models are the most basic and usually the place to start. $\endgroup$ Commented Jun 29, 2019 at 19:20
  • $\begingroup$ @HartoSaarinen OK, that is also my line of reasoning. But for that purpose are the authors then using the PP and ADF tests? $\endgroup$
    – WolfgangP
    Commented Jun 29, 2019 at 20:19
  • $\begingroup$ Well usually the OU process is assumed to be mean reverting. So I guess they want to verify that their AR1 process really is also. $\endgroup$ Commented Jun 30, 2019 at 8:33

1 Answer 1


For a stochastic process $X$ of the form:

$$ X_t = \alpha_0 + \psi X_{t-1} + \epsilon_t $$

where $\epsilon_t$ is white noise, the null hypothesis for the Dickey-Fuller test is

$$ H_0 : \psi = 1 $$


$$ H_A : \psi < 1 \, . $$

In case $H_0$ cannot be rejected, the process is non-stationary and a random walk with drift. In case $H_0$ is rejected, the process is stationary and an AR(1)-process and also a time-discrete O-U process.

  • $\begingroup$ You might want to correct your terminology "$H_{0}$ is accepted". $\endgroup$ Commented Jul 1, 2019 at 23:17

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.