While there has been much discussion about the promise of data-driven methods to study conflict, the reality is that successes have been very modest. A lot of effort has been spent either fitting parsimonious phenomenological models which aim to identify drivers for conflict or otherwise, big non-explanatory models which aim to learn complex relationshpis. The largest successes in conflict prediction (that we know of publicly) are the Uppsala project, the Predictive Heuristics project of Michael D Ward (source code available for both) which succeed at only predicting status quo: i.e. if a region is tradtionally peaceful, it will predict that, iif a region is traditionally volatile it will say so. Regions which switch into conflict from a peaceful duration seem beyond the grasp of these methods, and yet these are hidden from the metrics commonly used to assess accuracy because such events don't happen very often.

Nonstationariity, black swans, external factors all play a huge role, combined with unreliable and delayed data, and there's really no reason why we should expect to derive a clean explanatory mathematical relationship between conflict happening today and what happened in the past.

This motivates the question whether we can perform surveillance, and detect outbreaks, structural changes and sudden shifts in conflict dynamics. Such a method much be able to deal with non-stationarity, noisy, overdispersed data, and robust to reporting delays. Structural shifts needn't be simple shifts in mean, or fatality counts, but changes in structural dependence between conflict patterns in neighbouring regions.

The objective of this work is to develop the first steps for a generally applicable, open surveillance tool for conflict, able to interface with various disparate data sources, and able to, in an online fashion identify structural shifts in conflict data, making use of specific features identified by conflict experts.