I'm trying to test if a specific country's president's unexpected statements create a structural break in the exchange rate series (dollar).
I decided on what to do in what order but I'm having a hard time choosing models and tests. I'm not good at time series analysis and cannot decide trustfully. I know that the data has a unit root and is heteroskedastic intuitively, but after that, everything gets complicated. GARCH(1,1) looks like the best model, but I have a hard time motivating it in the report, also I do not know how to properly look for a variance break. Can anyone help?
Here is a part of my proposal:
First, I will test for a unit root, using the methodology provided by Phillips and Perron (1988). I choose this test because it is non-parametric, I do not have to select a lag level as in the ADF test. However, since this and the other methodologies I will use are based on asymptotic theory, I will have to use a bootstrap. Next, I will apply the White test to test for heteroskedasticity in the data. The reason I choose the White’s test over the Breusch-Pagan test is that it allows for nonlinear heteroskedasticity. I expect the time series data to have exponential heteroskedasticity, and will specify my model according to this assumption for now. After that, I will follow two parallel paths for detecting the breaks and breakpoints. While doing this, I will use joint estimation methods. 1. I test for an arbitrary number of breaks (unknown multiple breaks) for a specific time window with significant political activity and look for coinciding news. For this, I will use daily data. Normally, high frequency data proves to be problematic while testing for structural breaks, but at this step, I will eliminate this problem by chosing a small and specific time window. If I am to detect any breaks, I shall look for coinciding news and statements to describe the cause of the change. 2. I will chose important statements myself, accounting for the seperability of their effect on the exchange rate (whether there are other important events that could cause a structural break or not) and the element of surprise (to find a significant break that is not priced beforehand). Then I will use Chow test (for known break points) to determine whether these announcements cause the parameters to change or not. In both of these procedures, I will check for both mean and variance breaks, accounting also for the presence of trend and drift parameters.