Efficient Bayesian Inference for Switching State-Space Models using Discrete Particle Markov Chain Monte Carlo Methods

Switching state-space models (SSSM) are a very popular class of time series models that have found many applications in statistics, econometrics and advanced signal processing. Bayesian inference for these models typically relies on Markov chain Monte Carlo (MCMC) techniques. However, even sophisticated MCMC methods dedicated to SSSM can prove quite inefficient as they update potentially strongly correlated discrete-valued latent variables one-at-a-time (Carter and Kohn, 1996; Gerlach et al., 2000; Giordani and Kohn, 2008). Particle Markov chain Monte Carlo (PMCMC) methods are a recently developed class of MCMC algorithms which use particle filters to build efficient proposal distributions in high-dimensions (Andrieu et al., 2010). The existing PMCMC methods of Andrieu et al. (2010) are applicable to SSSM, but are restricted to employing standard particle filtering techniques. Yet, in the context of discrete-valued latent variables, specialised particle techniques have been developed which can outperform by up to an order of magnitude standard methods (Fearnhead, 1998; Fearnhead and Clifford, 2003; Fearnhead, 2004). In this paper we develop a novel class of PMCMC methods relying on these very efficient particle algorithms. We establish the theoretical validy of this new generic methodology referred to as discrete PMCMC and demonstrate it on a variety of examples including a multiple change-points model for well-log data and a model for U.S./U.K. exchange rate data. Discrete PMCMC algorithms are shown to outperform experimentally state-of-the-art MCMC techniques for a fixed computational complexity. Additionally they can be easily parallelized (Lee et al., 2010) which allows further substantial gains.

💡 Research Summary

Switching state‑space models (SSSMs) are a versatile class of time‑series models in which a continuous observation process is driven by a hidden discrete‑valued Markov chain. Bayesian inference for SSSMs requires sampling both the high‑dimensional discrete state sequence and any continuous parameters. Traditional MCMC schemes—Gibbs samplers that update each latent state one‑by‑one or Metropolis‑Hastings moves that modify small blocks—suffer from severe mixing problems because the discrete states are often strongly correlated across time. Consequently, convergence can be prohibitively slow, especially when the number of observations is large or when the model contains multiple change‑points.

Particle Markov chain Monte Carlo (PMCMC) methods, introduced by Andrieu, Doucet and Holenstein (2010), address this difficulty by embedding a particle filter inside an MCMC algorithm. The particle filter provides a high‑dimensional proposal for the whole latent trajectory, and the resulting “particle marginal Metropolis‑Hastings” (PMMH) algorithm targets the exact Bayesian posterior despite using an approximate likelihood estimator. However, the original PMCMC framework assumes a generic sequential Monte Carlo (SMC) algorithm, which is not optimal for discrete latent variables. In the discrete setting, specialized particle algorithms—first proposed by Fearnhead (1998) and later refined by Fearnhead & Clifford (2003) and Fearnhead (2004)—exploit the finite state space to construct exact or near‑exact proposals for the whole state sequence. These “discrete particle filters” avoid the usual particle degeneracy problem, dramatically reduce variance of the likelihood estimator, and can be up to an order of magnitude faster than standard SMC for the same number of particles.

The present paper introduces a new family of algorithms called discrete PMCMC. The key idea is to replace the generic particle filter inside PMCMC with the efficient discrete particle filter. Concretely, for a given set of model parameters, the discrete filter generates a weighted sample of complete state trajectories; these trajectories are then used as proposals in a Metropolis‑Hastings step that also updates the continuous parameters (either via a random‑walk proposal or a Gibbs step when conjugacy is available). The authors prove that the resulting Markov chain satisfies detailed balance with respect to the true posterior, i.e., the algorithm is exact despite using an approximate likelihood estimator. The proof hinges on the unbiasedness of the discrete filter’s likelihood estimate and on the fact that the filter’s proposal distribution coincides with the optimal conditional distribution of the state sequence given the parameters and the data, up to a normalising constant.

To demonstrate practical benefits, the authors apply discrete PMCMC to two distinct case studies.

1. Multiple change‑point model for well‑log data – The data consist of geological measurements along a borehole, where abrupt changes in rock properties correspond to discrete “regime” states. The model includes an unknown number of change‑points, their locations, and regime‑specific Gaussian means and variances. Using a standard Gibbs sampler, the authors observe very low effective sample sizes (ESS) for the change‑point locations even after tens of thousands of iterations. By contrast, discrete PMCMC with only 100 particles yields ESS values that are 5–10 times larger for the same CPU time, while posterior means and credible intervals match those obtained by the much more expensive Gibbs runs.

2. Regime‑switching model for US/UK exchange‑rate series – Here the hidden states represent different monetary‑policy regimes (e.g., high‑interest vs. low‑interest periods). The transition matrix is 4 × 4, and the observation equation is a stochastic volatility model. Standard PMMH with a bootstrap particle filter suffers from high variance in the likelihood estimator, leading to low acceptance rates (≈ 10 %). The discrete PMCMC algorithm, however, achieves acceptance rates above 40 % and reduces the Monte‑Carlo error of the log‑likelihood estimate by a factor of roughly three. Moreover, because each particle filter run can be executed independently, the authors parallelise the filter across 8 CPU cores, cutting wall‑clock time by more than 30 % relative to the serial implementation.

Beyond these empirical results, the paper discusses the natural parallelisability of discrete PMCMC. Following Lee, Andrieu and Doucet (2010), each particle filter can be run on a separate processing unit, and only the final likelihood estimate and the Metropolis‑Hastings acceptance decision require synchronization. This property makes the method well‑suited for modern multi‑core and GPU architectures, opening the door to real‑time Bayesian inference for large‑scale SSSMs.

In summary, the contributions of the paper are threefold:

• Methodological innovation – The authors define a generic discrete PMCMC framework that leverages the optimal proposal structure of discrete particle filters, thereby preserving exactness while dramatically improving efficiency.
• Theoretical validation – They provide rigorous proofs of unbiasedness, detailed balance, and convergence for the new algorithm, extending the PMCMC theory to a class of specialised particle filters.
• Empirical validation – Through two realistic applications, they show that discrete PMCMC outperforms state‑of‑the‑art MCMC and standard PMCMC in terms of effective sample size, acceptance rate, and computational speed, all while being straightforward to parallelise.

The paper concludes by suggesting future extensions, such as handling non‑linear observation models, incorporating adaptive particle numbers, and applying the approach to high‑dimensional multivariate SSSMs. Overall, the work represents a significant step forward in Bayesian computation for models with discrete latent dynamics, offering both theoretical rigor and practical performance gains.