Computational intelligence sequential Monte Carlos for recursive Bayesian estimation

Recursive Bayesian estimation using sequential Monte Carlos methods is a powerful numerical technique to understand latent dynamics of non-linear non-Gaussian dynamical systems. Classical sequential Monte Carlos suffer from weight degeneracy which is where the number of distinct particles collapse. Traditionally this is addressed by resampling, which effectively replaces high weight particles with many particles with high inter-particle correlation. Frequent resampling, however, leads to a lack of diversity amongst the particle set in a problem known as sample impoverishment. Traditional sequential Monte Carlo methods attempt to resolve this correlated problem however introduce further data processing issues leading to minimal to comparable performance improvements over the sequential Monte Carlo particle filter. A new method, the adaptive path particle filter, is proposed for recursive Bayesian estimation of non-linear non-Gaussian dynamical systems. Our method addresses the weight degeneracy and sample impoverishment problem by embedding a computational intelligence step of adaptive path switching between generations based on maximal likelihood as a fitness function. Preliminary tests on a scalar estimation problem with non-linear non-Gaussian dynamics and a non-stationary observation model and the traditional univariate stochastic volatility problem are presented. Building on these preliminary results, we evaluate our adaptive path particle filter on the stochastic volatility estimation problem. We calibrate the Heston stochastic volatility model employing a Markov chain Monte Carlo on six securities. Finally, we investigate the efficacy of sequential Monte Carlos for recursive Bayesian estimation of astrophysical time series. We posit latent dynamics for both regularized and irregular astrophysical time series, calibrating fifty-five quasar time series using the CAR(1) model. We find the adaptive path particle filter to statistically significantly outperform the standard sequential importance resampling particle filter, the Markov chain Monte Carlo particle filter and, upon Heston model estimation, the particle learning algorithm particle filter. In addition, from our quasar MCMC calibration we find the characteristic timescale τ to be first-order stable in contradiction to the literature though indicative of a unified underlying structure. We offer detailed analysis throughout, and conclude with a discussion and suggestions for future work.