She is doing research on cross listed stocks ( ADRs , GDRs) for valuation purpose she is using granger causality test through E views software, to compare the companies local and international stock market price. For that she need suggestion, of how many lags to be taken into consideration for the above test,
Hi Nuzhath,
I have run Johansen co-integration tests in STATA and am aware of Akaike Information Criterion (AIC) to automatically optimize for lags (this is specifically useful in panel data - accounts for heterogeneity of panels).
It would appear that this option is available for Johansen's cointegration test using Eviews. Perhaps see if EViews provides a similar option to AIC for the Granger causality test?
http://forums.eviews.com/viewtopic.php?f=18&t=1247
Also, during a cursory search of academic articles related to your specific problem, I did find this well-cited paper; this may be of use to your friend:
Lag-Length Selection and Tests of Granger Causality Between Money and Income
Daniel L. Thornton and Dallas S. Batten
Journal of Money, Credit and Banking
Vol. 17, No. 2 (May, 1985) , pp. 164-178
Published by: Ohio State University Press
Stable URL: http://www.jstor.org/stable/1992331
Hi,
I guess a two-step methodology should be used. Using a level VAR, try a lot of lags, and keep only those having spherical distrubances (Normal, no ARCH, no Autocorrelations....). Then among all this models, choose the one having the minimal AIC, or better the minimal AICc. Please note that using the AIC alone does not guaranty that residuals are correct, hence the two-step approach.
Best
Philippe
Hi,
i agree with Philippe;
also i would like to bring to your attention that other issues must be taken into concern:
AIC.AICc, & BIC focus on avoiding overfitting in a linear regression model; i think an important note is that all these metrics have a problematic constraints:
if lag (t-i) is included then all lags: (t-1) --> (t-i-1) are also included which is practically unrealistic; i.e. there is no reason to assume in advance that X(t-i) has greater influence than X(t-j) given that i < j;
i am working on such research (unpublished yet) and i have experiential evidence that Mutual Information is better to select only relevant lags. remember that the initial theory of (Granger 1969) emphasize that all variables must be included to conclude the existence of Causal relationship.
what i am trying to say is that (23. Thornton D. L. And Batten. D. S 1984) shows that the results of Granger-causality tests are extremely sensitive to the lag condition.
ref: 23. Thornton D. L. And Batten. D. S.: Lag Length Selection and Granger Causality. (1984) Working Paper Series 1984-001a Federal Reserve Bank Of St. Louis Available At: http://research.stlouisfed.org/wp/1984/1984-001.pdf
another issue must be taken into concern "asymmetric lag length selection": e.g. as for the response of Philippe you may want to check the following reference:
ref: Ozcicek, Omer; MacMillin, W. Douglas: Lag Length Selection in Vector Autoregressive Models: Symmetric and Asymmetric Lags. (1999) Applied economics- Abingdon : Routledge, ISSN 0003-6846, Vol. 31.1999, 4, p. 517-524
Additionally, from econometrics we found the results of Comincioli 1996 are experimental approval that it is normal to have different lag lengths in a Granger-causal test.
ref: Comincioli. B.: "The Stock Market As A Leading Indicator: An Application Of Granger Causality," THE PARK PLACE ECONOMIST: (1996) VOL. 4 available at: http://digitalcommons.iwu.edu/parkplace/vol4/iss1/13
hope that was helpful;
regards;
Hi Nuzhath,
The problem with E-view is that it performs the optimal-lag test using only the AIC criterion. If you want to check also for the contribution of the ammount of lags used, you have to test for it separately...
Try using the JMP platform (They have a free trial period), it shows all five common criteria at once and helps you determine optimal lags using an integrated approach. It also performs automatic sequential elimination, much like E-views, however lets you control which criterion it should follow.
In Eviews, open data as unrestricted VAR, then use View->LAg Structure->Lag Lenth Criteria. Opt for the Lag which has minimum of AIC, SC and HQ.
- Badges
- Science topic
- Similar topics
- Correction