Evaluating MCMC algorithms in Stochastic Volatility Fashions utilizing Simulation Primarily based Calibration
Authors: Benjamin Wee
Summary: Simulation Primarily based Calibration (SBC) is utilized to analyse two generally used, competing Markov chain Monte Carlo algorithms for estimating the posterior distribution of a stochastic volatility mannequin. Specifically, the bespoke ‘off-set combination approximation’ algorithm proposed by Kim, Shephard, and Chib (1998) is explored along with a Hamiltonian Monte Carlo algorithm applied by Stan. The SBC evaluation includes a simulation examine to evaluate whether or not every sampling algorithm has the capability to supply legitimate inference for the appropriately specified mannequin, whereas additionally characterising statistical effectivity by the efficient pattern measurement. Outcomes present that Stan’s No-U-Flip sampler, an implementation of Hamiltonian Monte Carlo, produces a well-calibrated posterior estimate whereas the celebrated off-set combination strategy is much less environment friendly and poorly calibrated, although mannequin parameterisation additionally performs a task. Limitations and restrictions of generality are mentioned.