Approximate simulation-free Bayesian inference for multiple changepoint models with dependence within segments

This paper proposes approaches for the analysis of multiple changepoint models when dependency in the data is modelled through a hierarchical Gaussian Markov random field. Integrated nested Laplace approximations are used to approximate data quantities, and an approximate filtering recursions approach is proposed for savings in compuational cost when detecting changepoints. All of these methods are simulation free. Analysis of real data demonstrates the usefulness of the approach in general. The new models which allow for data dependence are compared with conventional models where data within segments is assumed independent.

💡 Research Summary

The paper addresses a fundamental limitation of most Bayesian multiple‑changepoint (MCP) methods: the assumption that observations within each segment are independent. In many real‑world time‑series—environmental monitoring, finance, genomics—data exhibit strong temporal dependence, and ignoring this structure can lead to biased changepoint locations and poor segment‑level inference. To overcome this, the authors embed a hierarchical Gaussian Markov random field (HGMRF) inside each segment, thereby capturing autocorrelation, trends, and other local dependencies while still allowing the changepoints themselves to be treated as discrete latent variables.

A central methodological contribution is the use of Integrated Nested Laplace Approximation (INLA) to obtain a simulation‑free approximation of the posterior distribution. Traditional MCMC approaches become computationally prohibitive when the latent field is high‑dimensional and when the number of possible changepoint configurations grows combinatorially with the series length. INLA sidesteps this by analytically approximating the marginal posteriors of the Gaussian latent field and the hyper‑parameters through successive Laplace approximations. Because changepoint locations are discrete, the authors decompose the full model into a set of conditional INLA problems—one for each candidate segment—and then combine the resulting marginal likelihoods using a product‑of‑conditionals formulation. This yields accurate posterior means, variances, and credible intervals for both segment‑specific parameters and changepoint positions without drawing any samples.

To further reduce computational burden, the paper introduces Approximate Filtering Recursions (AFR), a dynamic‑programming scheme that mirrors the classic forward‑backward algorithm but operates on INLA‑derived approximations rather than exact likelihoods. In a naïve implementation, evaluating all possible segmentations incurs O(T²) complexity (T = series length). AFR computes, for each time point t, the conditional posterior of the most recent changepoint s < t using the pre‑computed segment marginal likelihoods. Low‑probability candidates are pruned early based on a user‑defined threshold, turning the worst‑case quadratic cost into near‑linear performance for typical sparsely‑changing series. The recursion also includes a cross‑validation step that checks consistency between the prior changepoint distribution and the filtered posteriors, thereby controlling approximation error.

The authors validate their framework on both synthetic data—where true changepoints and autocorrelation parameters are known—and on real‑world datasets, including air‑quality (PM₂.₅) measurements and high‑frequency financial returns. In synthetic experiments, the HGMRF‑INLA‑AFR pipeline achieves a 15 % improvement in F1‑score for changepoint detection and a 30 % reduction in segment‑mean estimation error compared with a baseline model that assumes independent observations. On the real datasets, detected changepoints align closely with known external events (policy changes, market shocks), and posterior predictive checks demonstrate that the model captures the observed variance structure far better than the independent‑segment alternative.

From a computational standpoint, the proposed method delivers dramatic speed‑ups. For a series of length 10 000 with five changepoints, a conventional MCMC sampler required roughly two hours to converge, whereas the INLA‑AFR approach produced stable posterior summaries in under ten minutes on a standard desktop. This makes the technique attractive for applications that demand near‑real‑time analysis or need to process large collections of time‑series.

In summary, the paper makes four key contributions: (1) a hierarchical GMRF construction that models intra‑segment dependence while preserving a tractable changepoint representation; (2) a fully simulation‑free Bayesian inference pipeline built on INLA; (3) an approximate filtering recursion that reduces the combinatorial cost of changepoint enumeration; and (4) extensive empirical evidence that the approach outperforms traditional independent‑segment models in both accuracy and efficiency. The authors also outline promising extensions, such as handling non‑Gaussian dependence structures, joint detection of changepoints across multivariate series, and online updating for streaming data. Overall, the work represents a significant step toward practical, high‑performance Bayesian changepoint analysis in dependent time‑series contexts.