Discussion of : Estimating the Historical and Future Probabilities of Large Terrorist Events

I congratulate Clauset and Woodard (2013) on a very interesting article. The authors analyze a global terrorism dataset with the aim of quantifying the probability of historical and future catastrophic terrorism events. Using power law, stretched exponential, and log-normal tail probability models for the severity of events (# deaths), the authors make a convincing argument that a 9/11 sized event is not an outlier amongst the catalog of terrorist events between 1968 and 2007. This study builds upon earlier work by Clauset, Young and Gleditsch (2007) that I also recommend for those interested in the statistical modeling of terrorism. While there is consensus amongst the models employed by Clauset and Woodard that 9/11 is not an outlier (p > .01), the estimates are accompanied by large confidence intervals on how likely a 9/11 sized event is. In Table 2, where the authors forecast the probability of a 9/11 sized event in 20122021, forecasted probabilities range from 0.04 to .94 depending on the model and the frequency of events over the time window. Here the uncertainty has less to do with the model specification and more to do with uncertainty in the frequency of events over the next decade. Terrorist events do not follow a stationary Poisson process and the intensity can fluctuate greatly over a several year period of time. The authors remark in their discussion that relaxing the stationarity assumptions and incorporating spatial and exogenous variables may help tighten the range of forecasted probabilities. I would add here that some progress has been made, in particular on modeling terrorist event time series as non-stationary point processes (see Porter and White (2012); Lewis et al. (2012); Zammit-Mangion et al. (2012); Mohler (2013); Raghavan, Galstyan and Tartakovsky (2012)). Terrorism event processes are history dependent and intensities exhibit correlations at timescales of weeks and months due to self-excitation (see Porter and White (2012); Lewis et al. (2012)) and