B1806
Title: Detecting abrupt changes in correlated time-series
Authors: Gaetano Romano - Lancaster University (United Kingdom) [presenting]
Guillem Rigaill - Universite Evry (France)
Paul Fearnhead - Lancaster University (United Kingdom)
Vincent Runge - Evry Paris-Saclay University (France)
Abstract: Change-Point analysis has been of major interest in recent times. Current algorithms for detecting changes in mean often struggle in the presence of auto-correlated noise, or in situations where the mean can vary locally between abrupt changes of interest. Default implementations of the algorithms in these scenarios will lead to detecting many spurious changes. This can be corrected for, but with a resulting loss of power. We develop principled statistical approaches to estimate changes under both these scenarios. These are based on maximising a penalised likelihood for appropriate models for the data. Estimating the change-points locations is non-trivial as it involves a solving a complicated, non-convex, optimisation problem. We show how to extend recent dynamic programming ideas to obtain exact solutions of this optimisation in (empirically) close to linear time in the number of observations. Our method is shown to out-perform alternative methods both on simulated and real-data.
