Statistics Help @ Talk Stats Forum - Applied Statistics
http://www.talkstats.com/
Statistics in finance, economics, engineering. Actuarial science. Econometrics. Operations research.enWed, 26 Apr 2017 12:44:36 GMTvBulletin60http://www.talkstats.com/images/misc/rss.pngStatistics Help @ Talk Stats Forum - Applied Statistics
http://www.talkstats.com/
Bayesian (structural) VAR model problems
http://www.talkstats.com/showthread.php/69340-Bayesian-(structural)-VAR-model-problems?goto=newpost
Tue, 25 Apr 2017 11:57:43 GMTHello, I will use a bayesian structural vector autoregressive model for a macroeconomic paper. It is very hard to find answers to many questions in...Hello, I will use a bayesian structural vector autoregressive model for a macroeconomic paper. It is very hard to find answers to many questions in academic papers which rarely presents all the steps in their thinking and model specification. I will use a minnesota prior where 3 hyper parameters need to be decided by me. In a normal non bayesian VAR model you can choose lag length by using an information criterion, you can also do autocorrelation, stability and normality tests etc. For example, autocorrelation behaves differently in a bayesian var model due to prior limitations on variables, thus limiting autocorrelation etc.

My question is, does anyone know where to find the criterion to set the "correct" lag length for bayesian var models? If this was a forecasting paper, I would naturally choose lag lengths and hyperparameter values to minimize some criterion (Root mean squared forecast error etc) but there is no clarity regarding the choice of lags for a structural bayesian var model. Also If anyone has a good idea regarding how to deal with eventual autocorrelation and how to comment it I would be thankful.

Thanks in advance.
]]>Applied Statisticsschteeke1http://www.talkstats.com/showthread.php/69340-Bayesian-(structural)-VAR-model-problemscopula-based VaR calculation in R
http://www.talkstats.com/showthread.php/69320-copula-based-VaR-calculation-in-R?goto=newpost
Sun, 23 Apr 2017 16:24:54 GMTI'm working on a value-at-risk calculation using copulas on different stock market indices. I know how to fit the copula, but I can't figure out how to apply the VaR approach in the next step. The concept of copulas is relatively new to me and has proven to be very challenging for an average master student..I defined 3 periods in which I want to investigate the evolution of the VaR over time. When running the code, R returns a value for the VaR. But when running the code for another time period, R gives the same value as the previous period..Am I overlooking/forgetting something? The code I provided shows the bivariate example of china and india using the normal copula. I plan to extent it with the t and clayton copula in a further stage.

Code:

library(copula)

cop_model = normalCopula(dim = 2)
m = pobs(as.matrix(cbind(CHINA_INDIA$CHINA.LOG[571:406],CHINA_INDIA$INDIA.LOG[571:406])))

#pseudo-observations
fit = fitCopula(cop_model, m, method = "ml")

coef(fit)

tau(normalCopula(param = coef(fit)))

cor(CHINA_INDIA$CHINA.LOG[571:406],CHINA_INDIA$INDIA.LOG[571:406], method = "kendall") #check whether correlation is more or less preserved

set.seed(1559)

u = rCopula(500, normalCopula(coef(fit), dim = 2)) #simulate some observations from the copula

cdf = pCopula(u, normalCopula(coef(fit), dim = 2) #construct cdf of the copula