A Toy Model Study for Long-Term Terror Event Time Series Prediction with CNN

Abstract

Followed by the 9/11 attacks in 2001 and the subsequent events, terrorism and other asymmetrical threat situations became increasingly important for security-related efforts of most western societies. In a similar period, the development of data gathering and analysis techniques especially using the methods of machine learning has made rapid progress. Aiming to utilize this development, this paper employs artificial neural networks for long-term time series prediction of terrorist event data. A major focus of the paper lies on the specific use of convolutional neural networks (CNNs) for this task, as well as the comparison to the performance of classical methods for (long-term) time series prediction. As the databases like Global Terrorism Database and Fraunhofer’s terrorist event database are not extensive enough to train a deep learning method, a simple toy model for the generation of time series data from one or more terrorist groups with defined properties is established. Metrics for comparison of the different approaches are collected and discussed, and a customized sliding-window metric is introduced. The study shows the principle applicability of CNNs for this task and offers constraints as well as possible extensions for future studies. Based on these results, continuation and further extension of data collection efforts and ML optimization techniques are encouraged.