Bayesian multiple change-point detection‎: ‎A methodological comparison of exact dynamic programming and nonparametric hidden Markov approaches

Detecting structural breaks is crucial for understanding time series dynamics‎, ‎particularly in highly volatile financial markets‎. ‎This study systematically compares two prominent Bayesian change-point methodologies‎: ‎an exact dynamic programming approach that eliminates Markov chain Monte Carlo sampling‎, ‎and a Dirichlet process hidden Markov model that flexibly infers latent regime counts and persistence‎. ‎Analyzing daily S&P ۵۰۰ returns spanning ۲۰۰۰ to ۲۰۲۴-a period encompassing major upheavals like the global financial crisis and the COVID-۱۹ pandemic-we evaluate both models across detection accuracy‎, ‎parameter estimation‎, ‎computational efficiency‎, ‎and predictive performance‎. ‎Despite fundamentally different assumptions‎, ‎the models exhibit striking concordance‎, ‎identifying major macroeconomic shifts with timing differences of mere days and parameter variations of under ۰.۰۱%‎. ‎However‎, ‎a clear trade-off emerges‎: ‎the exact method is approximately ۹۰ times faster‎, ‎whereas the Dirichlet process hidden Markov model provides marginally superior predictive accuracy‎. ‎These insights offer actionable guidelines for selecting the optimal framework based on computational and predictive priorities.

کلیدواژه ها:

‎Change-point detection ، ‎Dirichlet process ، ‎Dynamic programming