One of the focuses of interest for conflict research is crisis forecasting. While the approach often favored by the media and public approaches this challenge qualitatively with the help of pundits illustrating effects in a narrative form, quantitative models based on empirical data have been shown to be able to also provide valuable insights into multidimensional observations. For these quantitative models, Bayes networks perform well on this kind of data. Both approaches arguably fail to meaningfully include all relevant aspects as – expert knowledge is difficult to formalize over a complex multidimensional space and often limited to few variables (e.g. more A will lead to less B) – empirical data can only tell us about things that are easily measurable and can only show correlations (in contrast to causalities that would be important for forecasting) In this paper we will develop a method for combining empirical time series data with expert knowledge about causalities and “hidden variables” (nodes that belong to variables that are not directly observable), thereby bridging the gap between model design and fitting. We build a toolset to use operationalized knowledge to build and extend a Bayes network for conflict prediction and, model unobservable probability distributions. Based on expert input from political scientists and military analysts and empirical data from the UN, Worldbank and other openly available and established sources we use our toolset to build an early-warning-system combining data and expert beliefs and evaluate its predictive performance against recordings of past conflicts.    «

One of the focuses of interest for conflict research is crisis forecasting. While the approach often favored by the media and public approaches this challenge qualitatively with the help of pundits illustrating effects in a narrative form, quantitative models based on empirical data have been shown to be able to also provide valuable insights into multidimensional observations. For these quantitative models, Bayes networks perform well on this kind of data. Both approaches arguably...    »