Changepoint detection in non-exchangeable data

Abstract Changepoint models typically assume the data within each segment are independent and identically distributed conditional on some parameters that change across segments. This construction may be inadequate when subject to local correlation patterns, often resulting in many more changepoints fitted than preferable. article proposes a Bayesian changepoint model relaxes assumption of exchangeability The proposed supposes m -dependent for unknown $$m \geqslant 0$$ m ? 0 vary between segments, suitable detecting clear discontinuities different temporal correlations. approach is suited both continuous discrete data. A novel reversible jump Markov chain Monte Carlo algorithm sample from model; particular, detailed analysis parameter space exploited build proposals orders dependence. Two applications demonstrate benefits model: computer network monitoring via detection count data, segmentation financial time series.