What Is Terrorism Forecasting?

Terrorism forecasting is the application of statistical modeling, machine learning, and data analytics to predict the likelihood, location, timing, or type of future terrorist attacks. It belongs to the broader field of quantitative security analysis and draws on methods from time-series forecasting, natural language processing, spatial statistics, and behavioral science. The goal is to produce probabilistic estimates that allow security agencies, policymakers, and infrastructure operators to allocate protective resources before an attack occurs rather than only reacting afterward.

The field is empirically grounded in incident databases, the most widely used being the Global Terrorism Database (GTD), compiled by the National Consortium for the Study of Terrorism and Responses to Terrorism (START) at the University of Maryland. The GTD documents more than 200,000 terrorist incidents from 1970 onward, recording target type, attack method, geographic coordinates, group attribution, casualties, and dozens of additional attributes. This dataset provides the training ground for most published forecasting models.

Data Sources and Feature Engineering

Forecasting models require features that correlate with future attack probability. Incident-level historical data from the GTD provides the foundational time series, but models that rely solely on past attacks inherit the sparsity problem: many regions have few recorded events, limiting statistical power. Researchers have supplemented incident data with open-source intelligence feeds, including localized news text, social media content, and economic and governance indicators. Studies using localized news data, reported in PLOS ONE, demonstrate that incorporating regional news improves binary prediction of whether an attack will occur on a given date and in a given state compared to models using incident history alone. Spatial and temporal lag features, group-level activity indicators, and conflict escalation metrics further improve prediction.

Predictive Models

A range of machine learning architectures has been applied to terrorism forecasting. Classical approaches, including decision trees, random forests, support vector machines, and logistic regression, have been used to predict attack location or region, achieving accuracy between 75% and 90% on held-out test sets when applied to GTD records from 1970 to 2018. Deep learning methods treat the sequence of attacks as a temporal modeling problem: long short-term memory (LSTM) networks learn temporal dependencies across many years of incident records, while convolutional neural network (CNN) layers extract local feature patterns before passing representations to an LSTM for sequential modeling. A multilevel deep learning framework described in Scientific Reports reports binary classification accuracy above 96% and multi-class accuracy above 99% on held-out data when predicting attack occurrence and type. The Chatham House assessment of AI prediction and counterterrorism situates these technical results within a policy and operational context, noting the gap between in-sample model performance and real-world predictive utility.

Validation and Operational Limitations

Terrorism forecasting models face validation challenges that distinguish this domain from more stationary forecasting problems. Terrorist groups adapt their tactics in response to known surveillance and interdiction patterns, making historical associations between features and attacks less stable over time. Groups also evolve, merge, and disband, complicating the construction of group-level attribution models. Temporal autocorrelation in attack sequences violates the independence assumptions underlying many standard validation protocols; proper validation requires strict temporal splitting rather than random cross-validation. False-alarm costs and missed-detection costs are both high and asymmetric, making accuracy alone an incomplete performance metric. Published models rarely report calibrated probability scores or formal utility estimates that would allow decision-makers to translate predictions into resource allocation decisions.

Applications

Terrorism forecasting has applications in a range of security and analytical disciplines, including:

• Law enforcement resource allocation and patrol scheduling in high-risk areas
• Intelligence analysis support for prioritizing investigative leads
• Critical infrastructure risk assessment for hardening investment decisions
• Emergency preparedness planning for cities and transportation authorities
• Academic research on conflict dynamics and radicalization trajectories