Hi,
Currently the multivariate student and gaussian are the possible choices for rescor() in brms. Will there be support for other families? The aimed application is a financial VaR model where multiple risks are aggregated. The t-distribution can give VaR estimates that are hard to justify. Mixtures such as NIG generally fit observations better. Second question: Is there an alternative way to set up a VaR model in brms without rescor()?
EDIT:
The following seems to work, but convergence is worse:
library(MASS)
mu = c(1, 2, 4) # the mean of the variables V1, V2, V3
R = matrix(c(1, 0.7, 0.2, # correlation matrix (R)
0.7, 1, -0.5,
0.2, -0.5, 1), 3)
sigmas = c(0.5, 1.2, 2.3) # sd1=0.5, sd2=1.2, sd3=2.3
Sigma = diag(sigmas) %*% R %*% diag(sigmas) # Variance Covariance matrix
N=40
data = as.data.frame(mvrnorm(N, mu = mu, Sigma = Sigma)) # data simulation w/Sigma
data$time=1:N
fitO= brm(data = data,
family = gaussian,
bf(mvbind(V1,V2,V3) ~ 1+ (1|c|time))+set_rescor(FALSE),
chains=1,cores=8,iter = 9000, warmup= 1000,control = list(adapt_delta = 0.9, max_treedepth=15))
Any improvements to be suggested? Thanks!
Kind regards, Oyvind