In this section we compare the MCMC and importance sampling methods using artificially generated GARCH
data. We set the GARCH parameters to α = 0.05, β = 0.9 and ω = 0.1, and generated 3000 data. Fig.1 shows the
time series of the data.
We implement the MCMC method as in Ref.[13]. First we make a pilot run by the Metropolis algorithm to
estimate M and Σ of a multivariate Student’s t-distribution. The first 5000 Monte Carlo data by the Metropolis
algorithm are discarded as burn-in process. We then switch from the Metropolis algorithm to the MH algorithm.
During simulations we recalculate the values of M and Σ every 1000 Monte Carlo updates by using the accumulated
Monte Carlo data so far. Fig.2 shows the acceptance at the Metropolis test of the MH algorithm as a function
of every 1000 updates. At the beginning of the simulation the acceptance is low, which indicates that the values of
M and Σ are still not accurate enough. As we proceed simulations the acceptance increases rapidly and reaches a
plateau around 75%-80%. Final results are obtained by 200000 Monte Carlo data. The autocorrelation times are
calculated to be very small, 2τ ≈ 2, which means that the MH algorithm is a very efficient MCMC method for the
GARCH model[13].
The importance sampling algorithm also uses a multivariate Student’s t-distribution. We determine the values
of M and Σ for the multivariate Student’s t-distribution by a pilot run of the MH algorithm. Here we accumulate
5000 Monte Carlo data by the MH algorithm and calculate M and Σ of the multivariate Student’s t-distribution.