Sam Tickle

Analysing global terrorism incidence using a new high-dimensional changepoint detection method

Detecting changepoints in datasets with many variates is a challenge of increasing importance. In particular, the question of whether a change affects a large number of variates, or some small (and possibly very sparse) subset, is often critical in real-life applications. We'll discuss a new, efficient method that aims to distinguish between these two possibilities at a given changepoint. We call the resulting algorithm SUBSET, and demonstrate that it has good theoretical properties for the classical case of changes in mean under constant Gaussian noise. As SUBSET is a penalised likelihood-based method, it is applicable across a wide selection of possible changepoint problems in the parametric setting. For example, we can use SUBSET in situations with count data, such as in our modelling of the Global Terrorism Database. Specifically, we will investigate the evolving picture of the number of terrorist incidents per month for each global region since the 1970s, and attempt to reconcile our findings with potential historical causes.