B1390
Title: Regime changes and unsupervised learning
Authors: Marie Huskova - Charles University (Czech Republic)
Michal Pesta - Charles University (Czech Republic) [presenting]
Abstract: The purpose is to deal with a situation such that every occurrence of a phenomenon can cause several related events, and each event contributes to a different univariate counting process. Therefore, a collection of these dependent point processes forms a flexible multivariate counting process, where neither stationarity nor independence of interarrival times of the marginal processes is assumed. The main aim is to detect a structural break of some phenomena's occurrences over time, which means to test whether some (not necessarily all) intensities of the univariate counting processes are subject to change at some unknown time point. The asymptotic behaviour of the test statistic under the null hypothesis and the alternatives are investigated. Bootstrap add-on is proposed to overcome the computational curse of dimensionality and avoid nuisance parameters. The validity of the resampling technique is proved. A changepoint estimator is introduced as a by-product, and its consistency is provided. Multiple changepoints' detections are designed. The empirical properties are illustrated in a simulation study. The completely data-driven detection procedure is presented through an actuarial problem concerning claims from various insurance lines of business.
