Is Regression and Granger Causality Related?

[Cen Hua]

Hi Everyone!


Short Version:
Is it normal that "variable A" granger causes "variable B" but "variable A" is not a good regression independent variable for "variable B". So when you regress A as Independent and B as the dependent, A has really high P-value/low significance.


Long Version:
I am currently doing my dissertation investigating the relationship between Google search volume (search volume index = SVI) and stock prices, and I have two main objectives:

1. To investigate whether adding SVI into Fama-french's 3-factor model actually helps improve the model
(For those who are not familiar with the model: Dependent is the stock price premium or can also be seen as simply stock prices, Independent is some parameters not important for this question and the added Google SVI).
2. To investigate whether Google SVI granger causes stock prices movements (or the other way around)

So cut to the chase, I have analysed them using panel regression and found that Google SVI is a bad addition to the Fama-french's model (high SVI P-value and falling adjusted R squared when SVI is added). I.e. from this analysis, Google SVI is not an important variable to consider for predicting stock prices.

On the other hand, I found that Google SVI granger causes stock prices movements (Very low P-values from 0.17-0.47 between 2 periods lag to 7 periods lag).

I found that these results are counter-intuitive: if SVI granger causes the stock price movements it should also be a good variable addition to the Fama-French model, but it isn't. Is this normal, i.e. does not perform well in regression but does well in Granger test?

Thanks in advance!

[Sven-Kristjan Bormann]

Why do you think that p-values of 0.17 are low? Usually, only p-values less than 0.05 are considered to be low.

(Very low P-values from 0.17-0.47 between 2 periods lag to 7 periods lag).

[Phil Bromiley]

Remember, most folks on this listserv are not from your area so explain your problem in more general terms if you can. It would help if you more carefully follow the FAQ on asking questions – right Stata code in code delimiters, readable Stata output, and sample data using dataex.

For search volume to influence positively stock price, you would have to assume that most of the searches are prompted by positive intentions. One certainly could have a lot of searches prompted by bad news. It may be, and I vaguely remember seeing some research to this effect, that most such searches are aimed positively. You might also want to think about how this plays out theoretically – most of the three factor model attention is towards firm specific results and descriptions rather than the reaction of external parties.