The RATS Software Forum

Hi Tom,

I am currently using a multivariate GARCH BEKK Model to estimate volatility spillovers between tweets and stock market returns. If I estimate a bivariate GARCH model I get good results. However, I would like to estimate the model with more than two variables, at least more than 5. If I use 5 variables using the exact same code I use for the two (but with more variables of course), it does not converge. If I use 6 variables winRATS shuts down entirely, meaning the program closes itself without giving any output. Can you give me some help regarding this problem? I think the convergence issues might have to do with the fact that I only have about 180 observations, could that be the problem?
I cannot share the data, so I cannot share the program including the excel data file. Instead I will include parts of a copy of the code, sorry for the inconvenience. Hope you can still help.

Many thanks in advance, the copy is included below.

Kind regards,

Myrthe van Dieijen

OPEN DATA "C:\Users\K53SJ-SX247V\Desktop\Twitter_Data_v1.xlsx"
CALENDAR(7) 2009:10:1
DATA(FORMAT=XLSX,ORG=COLUMNS) 2009:10:01 2010:03:30 Abnormal_Returns Returns RETX Tw_Volume_Tweets Tw_Positive_Tweets $
Tw_Negative_Tweets Tw_Neutral_Tweets Tw_Positive_Negative_Ratio_Tweets Tw_Pos_Sub_Neg_Tweets Tw_Subjectivity_Tweets $
Tw_Retweets Tw_Retweet_Vol_Ratio Tw_Volume_TweetsSubRetweets Tw_Positive_Retweets Tw_Negative_Retweets $
Tw_Neutral_Retweets Tw_Positive_Negative_Ratio_Retweets Tw_Pos_Sub_Neg_ReTweets Tw_Subjectivity_Retweets $
Number_Blog_Posts Number_Forum_Posts Number_Print_News Number_Google_Search_Tickers Analyst_Estimates $
Ad_Spend Financial_Events New_Product_Launch Org_Events Comp_Volume_TweetsSubRetweets
@varlagselect(crit=sbc,lags=5)
# Returns Tw_Positive_Tweets Tw_Negative_Tweets Number_Blog_Posts Number_Forum_Posts

VAR Lag Selection
Lags SBC/BIC
0 12194.1363
1 11941.8945*
2 12034.7672
3 12124.2931
4 12226.7282
5 12311.0674

system(model=var1)
variables Returns Tw_Positive_Tweets Tw_Negative_Tweets Number_Blog_Posts Number_Forum_Posts
lags 1
det constant
end(system)
garch(p=1,q=1,model=var1,mv=bekk,rvectors=rd,hmatrices=hh,pmethod=simplex,piters=10,robusterrors)

MV-GARCH, BEKK - Estimation by BFGS
NO CONVERGENCE IN 200 ITERATIONS
LAST CRITERION WAS 0.0000007
With Heteroscedasticity/Misspecification Adjusted Standard Errors
Daily(7) Data From 2009:10:02 To 2010:03:30
Usable Observations 180
Log Likelihood -5903.2640

The default number of iterations for GARCH is 200, so you probably just need to increase that, since it appears to be very close to convergence at 200. Are your data reasonable candidates for a GARCH model? Four of the five series are count data. Are their values consistently large enough that they could reasonably be thought to be continuous? Also, you don't seem to be modeling those in logs. Is that intentional?