Probabilistic Early Warning for National Security Crises

Principal Investigator: Elizabeth Pate-Cornellsabeth Paté



The primary focus of the research for the CREATE project in the department of Management Science and Engineering at Stanford University has been to apply the tools of risk analysis to the ‘Indications & Warning’ function of intelligence analysis in order to improve the timeliness and accuracy of crisis warnings. We are building on past efforts to incorporate Bayesian assessment techniques into the analysis of intelligence, in particular efforts at the CIA in the 1970’s, but also academic efforts, including those of Paté-Cornell (2001), McLaughlin and Paté-Cornell (2005), Stech and Elsaesser (2007), and Pinker (2007). We address three shortcomings identified explicitly by CIA methodologists that are central to the nature of warning intelligence, but were not adequately addressed in subsequent academic literature. These issues are:   I.          Introducing dynamics into the Bayesian approach to both the evolution of the crisis and the probability assessment to account for the fact that signals are not received all at once, but rather, over time   II.         Incorporating the probability estimates into a decision to issue a warning that optimizes the satisfaction of the policy maker with consideration of lead time   III.       Capturing the dependences between intelligence signals without undue computational burden   Our approach utilizes a Partially Observable Markov Decision Process (POMDP) to model an underlying crisis scenario and integrate real-time Bayesian assessments of relevant intelligence data. It is a model that belongs to the larger class of Bayesian forecasting models explored by Harrison and Stevens (1976) in their seminal work. When solved using Bayesian assessment techniques assisted by computer software, the model yields an optimal decision whether to issue a warning (and at what level of urgency) or wait after each intelligence report is received and assessed. We also explore how to incorporate geolocational data regarding a crisis event into the POMDP in scenarios where uncertainty in the location of a potential crisis is a salient feature, as is the case when an analyst must assess local risk. Finally, we explore three other issues that are central to any formal model of crisis warning: effects of disaggregated decision making (multiple layers in the chain of command), means of effective scenario generation, and means of simplifying probability assessments (the latter being widely applicable to any model that employs Bayesian forecasting).   In the near term, we intend to apply this model to forecast both strategic-level and operational-level crises, possibly including nuclear weapons development and/or regime collapse (strategic level), and terrorist operations (operational level). The latter will exhibit uncertainty in the location of a crisis, not just the timing of a crisis. Eventually we expect the model to be incorporated into the set of tools in use at intelligence and law enforcement agencies to actively monitor the development of crisis scenarios in both a national and homeland security context, where analysts will use the tool to help them perform better their warning function.