Hello,
I would like to ask regarding my research. I am going to use the cointegration tests Trace and Eigen, but I need optimal lags for them. Then I want to use VAR or VECM model, but it will be based on results from cointegration tests. Should I use VARselect() function on data in levels or first order data I(0) as firstly I need to made the VAR model, but it requires data to be stationary, and my are stationary only when I made first differences. So please are my steps correct?:
1. VARselect() of optimal lag used on data in levels.
2. Cointegration test with optimal lag from 1. step but minus 1 as cointegration test should have one difference already in its tests.
3. VARselect() on differenced data and this lag used in VAR model or VECM based on cointegration results - but shouldn´t be the lag for VAR model and cointegration test the same? Because now I will have two different.
4. Test of VAR residuals on serial correlation via serial.test() function, do I have to use same lag length as in VAR model? Or can I use different lag length as used in VAR model. If I can use different lag length, than how to interpret or should I use for my VAR higher lag length that is without serial correlation? But always if my lag length is the same in serial correlation test and in VAR model, there is serial correlation, even if it is lag of order 20.
Thank you.
Are you using time series data?
Did you first test to see if your data is stationary or not? You can use an augmented Dickey Fuller test for this. If your data is stationary, even if it is time series, you might choose a different model, possibly even an ordinary least squares.
The VARselect is used to determine the optimal number of lags. You should run this this test on your time series object as is (not differenced). Print or otherwise view the results and then make a determination of how many lags to use on your var by the values provided for the different test statistics: AIC, HQ and SC (which is also know as BIC).
Remember that a time series that is not stationary may not be stationary in first differences. Even in first differences there may be the presence of a unit root. This is partly the role of choosing the correct number of lags.
How many variables do you have in your study?
Hello Daken,
yes, the first step was to test if the data are stationary as in cointegration test is requirement that data should be integrated at the same level.
I have 11 ETFs and 4 macro variables, each of the ETFs I want to test in pairs with each of the macro variables. I have 6 periods (2 ressions, 3 expansions and the whole period from 2005-2025).
I made the optimal lag selection for VAR model on the data in levels. But when I wanted to test the serial correlation I got different results in R and in Eviews. So I am not sure if my steps in R are okay. I used Breusch-Godfrey test for this in both. I also read that stability of the model must be tested and no one of the roots of the characteristic polynomial should be equal to 1, but I have some that are equal to 1 and a lot of them are almost one for example 0,9999. Is it a problem? I also made another stability test via stability() function type = "OLS CUSUM" and it is mostly okay. Is it maybe enought to use for the stability test OLS CUSUM and ignore the roots of the characteristic polynomial?
As I wasn't sure about the selection of optimal length, I went on the Toda Yamamoto principle. If you have some knowledge about this, could you advise if it is good to test it in R? Because again I got totally different results even if I choose the same length for the model and the same data in Eviews which is econometric software and in R. So I did my whole analysis in R via Toda-Yamamoto, but now I am not sure, which results are the right ones.
Here is part of my code with tests without data:
chosen_lag
What is the frequency of the data? Is it quarterly, monthly, daily?
Can you clarify the regression you are making? For example is one of the regressions ETF1 with macro1, macro2, macro3, and macro4?
Assuming you used the VARselect function in R, what was the output? How did you choose your lag length?
Feel free to email me directly at [email protected]
The VARselect function in R is typically used on stationary data, so it should be applied to first-differenced data if the original series are non-stationary; using it on data in levels that are not stationary can lead to misleading lag length selection and unreliable VAR estimates.
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series