Granger Causality Analysis and testing assumptions

[Truike11]

I have a question regarding the Granger Causality Analysis.

I'm conducting a research about the influence of Twitter on the stock market. So I gathered tweets from 98 different stocks and matched them with the 98 stock quotes. I'm using intraday (by the hour) quotes. I performed the Augmented Dickey Fuller (levels) for all the tweets and the returns of the quotes and afterwards estimated the lags using the VAR and the lag length criteria. Now my question is, do I have to check for normality and heteroscedasticity? I tried it already but cannot find what the results exactly mean (as shown below). The results below is from 1 time serie of tweets and the matching returns. If I do have to test for normality and heteroscedasticity before running the GCA, could someone explain how to interpret the tables?

Thanks in advance for your time, very much appreciated!

VAR Residual Heteroskedasticity Tests: No Cross Terms (only levels and squares)
Date: 11/28/13 Time: 13:36
Sample: 10/07/2013 09:00 10/18/2013 16:00
Included observations: 75


Joint test:

Chi-sq df Prob.

20.55626 24 0.6647


Individual components:

Dependent R-squared F(8,66) Prob. Chi-sq(8) Prob.

res1*res1 0.059771 0.524456 0.8340 4.482806 0.8112
res2*res2 0.101858 0.935628 0.4936 7.639339 0.4695
res2*res1 0.062075 0.546012 0.8175 4.655617 0.7937


VAR Residual Normality Tests
Orthogonalization: Cholesky (Lutkepohl)
Null Hypothesis: residuals are multivariate normal
Date: 11/28/13 Time: 13:42
Sample: 10/07/2013 09:00 10/18/2013 16:00
Included observations: 75


Component Skewness Chi-sq df Prob.

1 8.049514 809.9335 1 0.0000
2 -0.039984 0.019984 1 0.8876

Joint 809.9535 2 0.0000


Component Kurtosis Chi-sq df Prob.

1 68.14146 13260.66 1 0.0000
2 3.087196 0.023760 1 0.8775

Joint 13260.68 2 0.0000


Component Jarque-Bera df Prob.

1 14070.59 2 0.0000
2 0.043744 2 0.9784

Joint 14070.63 4 0.0000

[CharlieEVIEWS]

Whatre the null hypotheses of these tests? What're the p-values in your output tables? What do the p-values suggest about the null hypotheses? Answering these questions will give you the answer to your question.

[Truike11]

Sorry, the models are not really my expertise but I have to use them. I understand that I have to look at the p-values. But especially for the normality test, they are so low that I cannot figure out how to make the variables normally distributed. Moreover, to remove the serial correlation, Brooks points out that I have to include lags but I already include lags when I use the Granger Causality Analysis so I cannot figure either how to remove it.

[CharlieEVIEWS]

This post (and subsequent comments, of which there are many) by Professor Dave Giles may be of assistance to you:

http://davegiles.blogspot.ca/2011/04/te ... ality.html

While it specifically investigates the case of non-stationary, the parts about normality and serial correlation will still be of relevance even if your data are stationary (I think - happy to be corrected.)

[Truike11]

Thanks!