Well? Does all variables in a VAR/VEC need to be normally distributed, or only the target variable? It is very hard to get all of them to meet criteria of normality without deleting too many outliers.
Lars:
None of them needs to be normally distributed. Not even the error term for many purposes. See:
http://cadmus.eui.eu/bitstream/handle/1814/19354/ECO_2011_30.pdf?sequence=1; page 13.
Indeed, just departures from normality in terms of skewness (not kurtosis) entails some bias in cointegration tests. See http://web.calstatela.edu/faculty/klai/KLPaper/OBES93Au.pdf; page 322.
Best regards.
Excuse-me, the first link wasn't appropiate.
It is http://cadmus.eui.eu/bitstream/handle/1814/19354/ECO_2011_30.pdf?sequence=1
Hope it helps
Thanks! However, I know that the trace statistic from the Johansen cointegration test is robust both to kurtosis and to skewness, but that was not really my question. I'm wondering about the VAR with error correction.
Lars...
Excuse-me. I'm not sure about your question. You wrote: "Does all variables in a VAR/VEC need to be normally distributed, or only the target variable?"...... My answer was: "None of them needs to be normally distributed". This is because I read your question literally. I'm Sorry.
If your question was about VECM residuals (not variables), a simple answer could be ... No, you just need some ioint normality Null Hypothesis not to be rejected (you can test that with many different tests as http://www.stata.com/manuals13/tsvecnorm.pdf).
But once again, not even the VECM residuals must to be normally distributed for some specific purposes (e.g. impulse-response functions. Please see some comments in-here:http://davegiles.blogspot.com.ar/2012/03/overview-of-var-modelling.html)
Best regards.
- Badges
- Science topic