Stochastic volatility (with a custom distribution)

library(nimble, warn.conflicts = FALSE)

## nimble version 0.13.0 is loaded.
## For more information on NIMBLE and a User Manual,
## please visit https://R-nimble.org.
## 
## Note for advanced users who have written their own MCMC samplers:
##   As of version 0.13.0, NIMBLE's protocol for handling posterior
##   predictive nodes has changed in a way that could affect user-defined
##   samplers in some situations. Please see Section 15.5.1 of the User Manual.

stochvol_installed <- require(stochvol, quietly = TRUE)

Stochastic volatility models are often used for time series of log returns of financial assets. The main idea is to model volatility (standard deviation of log returns) as an unobserved autoregressive process.

This example shows coding of a stochastic volatility model in nimble. See the particle filter and particle MCMC (PMCMC) examples for demonstrations of these methods with this model. We follow the example model and data described in the vignette (“Dealing with Stochastic Volatility in Time Series Using the R Package stochvol”) of the stochvol package (Kastner 2016, Kaster and Hosszejni 2019). We reparameterize relative to the stochvol example as noted below.

Code for this model is:

stochVolCode <- nimbleCode({
  x[1] ~ dnorm(0, sd = sigma / sqrt(1-phi*phi))
  y[1] ~ dnorm(0, sd = beta * exp(0.5 * x[1]))
  for (t in 2:T){
    x[t] ~ dnorm(phi * x[t-1], sd = sigma)
    y[t] ~ dnorm(0, sd = beta * exp(0.5 * x[t]))
  }
  phi <- 2 * phiStar - 1
  phiStar ~ dbeta(20, 1.1)
  logsigma2 ~ dgammalog(shape = 0.5, rate = 1/(2*0.1)) ## This is Omega
  sigma <- exp(0.5*logsigma2)
  mu ~ dnorm(-10, sd = 1) ## It matters whether data are converted to % or not.
  beta <- exp(0.5*mu)
})

In this model:

y[t] is the daily log return, i.e. log(exchange rate at time t / exchange rate at time t-1). These are the observed data.
x[t] is the latent state related volalitility that undergoes a linear autoregressive stochastic process with autocorrelation phi and noise standard deviation sigma.
The distribution of x[1] is the stationary distribution of the x[t] stochastic process.
beta is a constant for the volatility.
Standard deviation of observations is beta * exp(0.5 * x[t]).
The strongly informative prior on phi is parameterized in terms of phiStar following the stochvol vignette.
The priors for sigma and beta also follow the stochvol vignette.
The prior for logsigma2 requires a custom probability density calculation, provided below, such that exp(logsigma2) follows a gamma distribution with shape and rate as given in the code.

The probability density needed for logsigma2 is given by the following custom distributions:

dgammalog <- nimbleFunction(
  run = function(x = double(), shape = double(),
                 rate = double(),log = integer(0, default = 0)) {
    logProb <- shape * log(rate) + shape * x - rate * exp(x) - lgamma(shape)
    if(log) return(logProb)
    else return(exp(logProb))
    returnType(double())
  }
)

rgammalog <- nimbleFunction(
  run = function(n = integer(),
                 shape = double(), rate = double()) {
    xg <- rgamma(1, shape = shape, rate = rate)
    return(log(xg))
    returnType(double())
  }
)

The “r” function can be skipped if it will be not be needed, but we include it here for completeness.

Again following the stochvol vignette, we use as data exchange rates for the Euro (EUR) quoted in U.S. Dollars (USD). Here we choose to start after January 1st, 2010, and continue until the end of the time-series, 582 days after that.

Set up the data

data('exrates')
y <- logret(exrates$USD[exrates$date > '2010-01-01'], demean = TRUE)

Build the model

stochVolModel <- nimbleModel(code = stochVolCode,
                             constants = list(T = length(y)), data = list(y = y),
                             inits = list(mu = -10, phiStar = .99,
                                          logsigma2 = log(.004)))

## Defining model

##   [Note] Registering 'dgammalog' as a distribution based on its use in BUGS code. If you make changes to the nimbleFunctions for the distribution, you must call 'deregisterDistributions' before using the distribution in BUGS code for those changes to take effect.

## Building model

## Setting data and initial values

## Running calculate on model
##   [Note] Any error reports that follow may simply reflect missing values in model variables.

## Checking model sizes and dimensions

##   [Note] This model is not fully initialized. This is not an error.
##          To see which variables are not initialized, use model$initializeInfo().
##          For more information on model initialization, see help(modelInitialization).

Compile the model

CstochVolModel <- compileNimble(stochVolModel)

## Compiling
##   [Note] This may take a minute.
##   [Note] Use 'showCompilerOutput = TRUE' to see C++ compilation details.

See further examples for using this model.