Learning Made Easy: A Marginalized Resample-Move Approach

• Activity Number: 134
• Type: Contributed
• Date/Time: Monday, August 1, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Business and Economic Statistics Section
• Author(s): Junye Li*+
• Companies: ESSEC Business School
• Keywords: State-space models ; Particle filters ; Parameter learning ; State filtering ; Resample-move ; Stochastic volatility

Abstract

Parameter learning in state-space models, especially in dynamic asset pricing models, is practically difficult. This paper proposes a simulation-based parameter learning method in the general state space models. First, the approach breaks up the interdependence of the hidden states and the static parameters by marginalizing out the states using a particle filter. Second, it proposes a Bayesian resample-move approach to this marginalized system. This marginalized resample-move is exact in the sense that for any fixed number of M particles used to in hidden states, it delivers sequential samples from the posterior distributions as the number of particles over the fixed parameters, N, goes to infinity. Simulation studies show that our learning method can deliver the same posterior outputs as standard MCMC methods, both in a linear Gaussian model and a nonlinear non-Gaussian one. More importantly, it provides posterior quantities necessary for full sequential inference and recursive model monitoring. Furthermore, the methodology is generic and needs little design effort. The algorithm is also implemented on real data for a stochastic volatility model and a credit risk model.