Can we predict eruptions of violence? Statistics and the future of conflict early warning

To my great surprise, the answer might be “yes”.

There has been a surge of interest in conflict early warning. More like a frenzy.

Ushahidi has become one of the most-talked-about NGOs of the decade. They developed a platform for crowdsourcing real-time crisis information (like incidents of violence) from  web and text message.

The LRA crisis tracker does something similar, focused on the movements and actions of the marauding Lord’s Resistance Army.

I’ve been skeptical of each, and haven’t decided whether they are advocacy tools or actual tools. But I think the answer is both, and I can see the value for planning and rapid response. Both are important experiments, possibly over-hyped, possibly not.

It would be nice, of course, if early warnings came, well, earlier. What if we could forecast where mass violence was likely to break out?

Police have been trying this out in rich cities with fancy statistics, to mixed results at best. My view has been that mass violence is fairly idiosyncratic: many places have the potential, but picking the most likely places it will erupt is fairly futile.

A project in Liberia provided an opportunity to see. From 2009 through 2010 my coauthers and I studied whether the government and UN could change the norms and institutions around conflict and dispute resolution in 250 towns and villages. With virtually no formal security or justice sector, the informal system delivered justice, and the aim was to improve it. (Policy report here, and draft academic paper, “Institution building at the local level”–ready in the next couple of weeks, I hope.)

Following a panel of 250 violence prone villages, and 10,000 surveys, sparked an idea: Could we use the 2009 data to predict mass violence–communal killings, ethnic conflicts, or mob violence–in 2010? I was almost sure the answer would be no.

As usual, nearly every one of my field experiments and studies prove my expectations wrong. I must be the worst qualitative observer on the planet.

Here is our answer (the lite policy memo version–the detailed academic paper will take much more time). Even without the best 2009 data (we hadn’t planned this at the time) and even with little time-varying data (we only had two rounds) we do unexpectedly well.

Some highlights:

  • We correctly predict up to 75% of all conflicts two years later
  • Simpler model with fewer factors do better than more interactive models with many factors
  • We can train the models not only to maximize accuracy, but to minimize “false negatives”–the costly cases where you predict peace instead of violence
  • We can identify 40 to 70 percent of all incidents (“true positives”), with three to five false alarms (or “false positives”) for each correctly predicted incident


Some caveats: these are not the ideal data or sample, we are not skillful statistical forecasters (we are learning). I am looking forward to May, when a conference of forecasting luminaries are going to shine harsh light on our foibles.

I can imagine a Ushahidi-like system of crowdsourcing could be harnessed to give early, early warning if one designed a system that pulled in consistent and complete data, ideally at regular frequency. This would be interesting to develop and test. You know, with all that free research time I have. We will see.

I don’t think this would work for the LRA crisis tracker, or its imitators. There is a difference between predicting the actions of many people perpetrating many like acts in many places, and predicting the actions of an individual or small force. Game theory and people like Bruce Bueno de Mesquita may be a better guide there.

Yes, I realize I am beginning to sound like Hari Seldon.

Reader pointers to similar forecasting and prediction efforts welcome. I am almost certainly ignorant of some of the interesting things going on.