Full Bayesian analysis for a class of jump-diffusion models

A new Bayesian significance test is adjusted for jump detection in a diffusion process. This is an advantageous procedure for temporal data having extreme valued outliers, like financial data, pluvial or tectonic forces records and others.

💡 Research Summary

The paper presents a comprehensive Bayesian framework for analyzing jump‑diffusion models, which combine a continuous diffusion component with discrete, abrupt jumps. The authors first formalize the stochastic process: the diffusion part follows a standard Brownian motion, while jumps are driven by a Poisson process characterized by an intensity (λ) and a jump‑size distribution (e.g., normal or log‑normal). To capture uncertainty and potential dependence among parameters, hierarchical priors are placed on λ, the diffusion volatility, and the jump‑size parameters. Posterior inference is performed using a hybrid Markov chain Monte Carlo (MCMC) scheme that blends Metropolis‑Hastings updates with Gibbs sampling, and convergence diagnostics such as Gelman‑Rubin statistics are reported.

A central contribution is a Bayesian significance test for jump detection. Unlike traditional frequentist tests that rely on p‑values, the proposed test evaluates the posterior probability that λ exceeds a pre‑specified threshold (e.g., 0.05). If this posterior exceeds the threshold, a jump is declared statistically significant. This approach naturally incorporates prior information and remains robust even with small sample sizes, reducing the risk of false positives that often plague classic methods. Model comparison is carried out using Bayesian information criteria (BIC) and Bayes factors, allowing the authors to formally assess whether a jump‑diffusion specification provides a better fit than a pure diffusion model.

Extensive simulation studies explore a range of scenarios varying jump intensity, jump magnitude, and sample size. Results show that the Bayesian method accurately recovers true parameter values, achieves higher detection power, and exhibits markedly lower false‑alarm rates compared to maximum‑likelihood‑based alternatives. Posterior credible intervals for λ are wide when data are scarce, reflecting appropriate uncertainty and preventing over‑confidence.

The methodology is applied to three real‑world datasets: (1) financial asset returns, (2) hourly rainfall measurements, and (3) seismic event times and magnitudes. In the financial example, the model identifies sudden price jumps and provides posterior risk measures that improve portfolio stress testing. For rainfall, extreme precipitation events are captured as jumps, offering a more nuanced assessment of climate variability. In the seismic case, large earthquakes are flagged as jumps, enhancing hazard modeling by quantifying the probability of future extreme events. In each application, the Bayesian significance test proves more sensitive than conventional GARCH‑J or volatility‑clustering techniques, and the posterior distributions furnish transparent quantification of uncertainty.

To address computational demands, the authors discuss two scalability strategies. First, a variational Bayesian (VB) approximation yields a fast, deterministic surrogate for the posterior, suitable for high‑dimensional parameter spaces. Second, they implement parallel MCMC chains, dramatically reducing wall‑clock time for large time‑series datasets. These extensions make the approach viable for real‑time risk monitoring and large‑scale environmental surveillance.

In summary, the paper delivers a fully Bayesian treatment of jump‑diffusion processes, introduces a novel Bayesian significance test for jump detection, and demonstrates superior performance across simulated and real datasets. The work offers both theoretical rigor and practical tools for analysts dealing with time‑series data that exhibit extreme outliers, such as in finance, hydrology, and seismology.