A0898
Title: Mutually dependent Bernoulli processes for multivariate change-point detection
Authors: Carson McKee - Kings College London (United Kingdom) [presenting]
Maria Kalli - Kings College London (United Kingdom)
Abstract: Financial and economic time series exhibit sharp structural breaks, or change-points, driven by events such as recessions or financial market crashes. In the multivariate setting, these changes may occur asynchronously across series. It is often the case that the occurrence of a change-point in one series affects the probability of a subsequent change occurring in another. This is observed, for example, in financial contagion, when turmoil spreads from one country to another. A multivariate change-point prior is developed, which explicitly models this dependence using discrete time and mutually dependent point processes. Under this prior, the probability of a change-point occurring in a series at a given time is dependent on recent changes in other series. Thus, the model allows for both cross-sectional and temporal dependence in the change-point probabilities. Then, conditional on the change-point locations, the data in each segment is assumed to be independent of the data in other segments. Obtaining posterior estimates under this model is non-trivial. A blocked Gibbs sampler and a particle Gibbs sampler are developed for use in high dimensions. The model is demonstrated on simulated and real datasets and is shown to uncover latent dependencies linking change-points in different series.
