ABSTRACT
There has been a continued interest in modeling the activity profile of terrorist groups over the last few decades. More recent developments in this area have been along two directions. The first framework leverages a self-exciting hurdle model, popularized in seismology and gang warfare modeling, for terrorist activity. The second framework builds a hidden Markov modeling framework to capture terrorist group dynamics. The focus of this chapter is on a comparative analysis of these two frameworks in terms of their explanatory and predictive powers. Specific attention is paid to the inferencing capability of the hidden Markov modeling framework for the early detection of spurts and downfalls. In this direction, a parametric approach that performs retrospective state classification via the Viterbi algorithm, a non-parametric approach performing an exponentially weighted moving average filtering, and a non-parametric approach motivated by majorization theory are proposed. While the first approach leads to good performance, it is not practically amenable due to model learning latencies. The second approach avoids these latencies, but it can only detect major activity spurts. On the other hand, the third approach combines the good properties of both these approaches with a good performance in spurt detection.
