Hi Tom and everyone,
I have 12 exchange rates variables with log form, which are all I(1) process. Thus I want to do the rolling window regression for Cointegrated VAR model, namely first use Johansen(1991) test to determine the number of cointegration vectors, if the number is zero, use one order VAR model to estimate and variance decomposition; if not, which means there's a cointergration relationship among variables, do reduced rank regression and use VECM model to estimate and variance decomposition. This method is based on Phillips(1998), Impulse Response and Forecast Error Variance Asymptotics in Nonstationary VARs, Journal of Econometrics, 83(1-2):21-56.
I want to use this Cointegrated VAR model in rolling window regression, but I don't know how to write the Rats code, especially the optimal lag order of VAR and number of cointegration vectors are determined dynamically in rolling window regression. So how can I do this?
Thank you in advance.
rolling window regression for Cointegrated VAR model
-
Alice Zhang
- Posts: 3
- Joined: Thu Jul 20, 2017 2:30 am
Re: rolling window regression for Cointegrated VAR model
Rolling sample analysis can be set up relatively easily with the rolling analysis dialog. Before you do that, however, you want to make sure you've figured out how to do a full sample analysis, because what you put inside the loop is the same thing with restricted range.
Trying to pick the cointegration rank using rolling sample statistics seems like a really bad idea. If the test for r=0 vs r=1 is "significant" for 10 periods and "insignificant" otherwise, do you really think the series are cointegrated during those 10 periods and nowhere else? What would that even mean? (You might want to read the comments in the example on rolling causality tests on how rolling sample analysis can lead one to overinterpret sampling error variations). CATS presents recursive tests for cointegrating rank, but does them by showing the time series of test statistics. Note, however, that those are done primarily as a diagnostic for the decision on rank for the full-sample estimates and use increasing samples, rather than rolling windows.
Trying to pick the cointegration rank using rolling sample statistics seems like a really bad idea. If the test for r=0 vs r=1 is "significant" for 10 periods and "insignificant" otherwise, do you really think the series are cointegrated during those 10 periods and nowhere else? What would that even mean? (You might want to read the comments in the example on rolling causality tests on how rolling sample analysis can lead one to overinterpret sampling error variations). CATS presents recursive tests for cointegrating rank, but does them by showing the time series of test statistics. Note, however, that those are done primarily as a diagnostic for the decision on rank for the full-sample estimates and use increasing samples, rather than rolling windows.