README.md

bExtremes

R package for extreme risk estimation using extreme value theory and Bayesian inference.

Automatic threshold weighting is done with body-tail models. Implemented body distributions include "GH" and "GP". The latter is recommended.

A fixed-threshold GP model is also included for comparisons (see functions createGPData and createGPModel).

Usage

loadNamespace('devtools')
devtools::dev_mode(on=TRUE)

#Requires LaplacesDemon package from github repo: samedii/LaplacesDemon
attempt <- try(loadNamespace('LaplacesDemon'))
if(inherits(attempt, 'try-error')) {
    devtools::install_github('samedii/LaplacesDemon')
}

attempt <- try(library(bExtremes))
if(inherits(attempt, 'try-error')) {
    devtools::install_github('samedii/bExtremes')
}
library(bExtremes)

#cat('Getting data...\n')
#orgData <- fetchExampleStockData('BAC') #function unavailable, replace with your own

#Transforming and flipping data (because right tail is modelled)
#transformedData <- -transformStockData(orgData)

#cat('Filtering volatility...\n')
#transformedData <- filterVolatility(transformedData)

cat('Simulating data...\n')
transformedData <- simulateGHGP()

cat('Creating prior...\n')
#Expert prior (Guessing from historical data)
sortedData <- sort(transformedData)
##Median for VaR at 0.01
m <- sortedData[ceiling(length(sortedData)*(1-0.01))]
##90% quantile for VaR 0.01
q <- m*1.2

##Median for difference between VaR at 0.001 and 0.01
mDiff <- sortedData[ceiling(length(sortedData)*(1-0.001))]-m
##90% quantile of above
qDiff <- mDiff*1.2

prior <- createExpertPrior(m,q,mDiff,qDiff,transformedData)

#Uninformed prior (reference prior)
prior <- referencePrior

#Calculate VaR and ES for confidence level p (preferably 0.99 or greater):
p <- 0.99

cat('Setting up model...\n')
data <- createData(transformedData,p,dist='GP')
model <- createModel(data,prior)

cat('Generating IV...\n')
ivAndCov <- generateIV(model,data,n=500) #It can be helpful to increase n

cat('MCMC process...\n')
fit <- fitMCMC(model,data,ivAndCov)

#Information about final MCMC run and results
LaplacesDemon::Consort(fit)

#Summary of posterior samples
fit$Summary2

#Samples
head(fit$Posterior2) #parameters
head(fit$Monitor) #monitored variables (including VaR and ES)

#Example plots
hist(fit$Monitor[,7],breaks=100)
LaplacesDemon::joint.density.plot(fit$Monitor[,7],fit$Monitor[,8])
LaplacesDemon::joint.pr.plot(fit$Posterior2[,3],fit$Monitor[,7])