Statistical forecasting isn’t perfect but it’s always getting better. The international community should sit up and listen because it could help save lives.
In March 2012, junior officers stage a coup in Mali, throwing the country into disarray. A year later, rebels oust the government of the Central African Republic (CAR), paving the way for widespread violence that has made refugees out of a quarter of the country’s population. And at the end of the year 2013, an internal political conflict in South Sudan’s governing party and army escalates into a full-scale civil war, killing ten thousand or more.
These conflicts differ widely in almost every aspect, apart from the sense of surprise and helplessness that they instilled in the international community. Mali was lauded as a democratic role model before some soldiers took power almost by accident. The French government, for decades the kingmaker of the CAR, confessed to being taken blindsided by the speed and viciousness with which the conflict escalated. And in South Sudan, the regional organisation IGAD struggled to respond to the conflict, finding themselves unprepared and at odds over how exactly to proceed.
In all three cases the surprise greatly limited the influence of the international community, which if better prepared could not only have intervened earlier and more effectively but could perhaps even have taken pre-emptive measures. This unpreparedness was even more of a shame because in all three cases, the outbreak of conflict had been predicted by statistical models.
Jay Ulfelder is one of the leading academics in the field of conflict forecasting. Since 2012, he has published a yearly coup forecast and works on predicting state-led mass killings for the US Center for the Prevention of Genocide.
Of course, not all top-ranking countries experienced coup attempts, but that too was to be expected, according to Ulfelder. “The probabilities of these things happening in a particular calendar year are generally very low,” he says, “but they can certainly be high-impact events, so they might warrant attention nonetheless.”
Ulfelder uses an array of data for each country in his coup and state-led mass killing forecasting models, including economic growth and income data, as well as indices for the state of democratic rule and the time since the last coup attempt. He then calculates the risk of the occurrence of conflicts based on the circumstances under which conflict occurred in the past: if a certain characteristic (e.g. poverty) has been associated with conflict in the past, countries that continue to display that characteristic will have a higher likelihood of experiencing violence again.
In a sense, these statistical models “learn” from the past, says Ulfelder, but this is also one of their main weaknesses: “[no model] would have shown Egypt as especially high risk in 2011,” he points out. “The models ‘learn’ from the history they’re shown, and in the past few decades, most coups have occurred in poor countries with competitive authoritarian or illiberal democratic political regimes, not middle-income autocracies [like Egypt]. It’ll be interesting to see how the models and forecasts change once I include data from the past few years in the estimation process.”
According to Ulfelder, one the main impediments to more accurate forecasts is the timeliness of the data available. “I need data from 2013 to produce 2014 forecasts. Unfortunately, many of the sources for the measures used in these models won’t publish their 2013 data for at least a few more months,” he writes in the blog post introducing his latest forecast. This means that in his 2014 forecasts, for example, he is largely using data from 2012.
Especially given these shortcomings, Ulfelder believes the outcome of his 2014 forecast is quite impressive. “I think the fact all of the cases in which we think we saw onsets of state-led mass killing in 2013 – Egypt, CAR, and possibly Nigeria and South Sudan – show up in the top 20, and mostly in the top 10 using data that’s almost all from 2012 indicates that the forecasting process works reasonably well,” he says.
Getting better data
One dataset that could do much to improve the timeliness of forecasts is the Global Database of Events, Language and Tone (GDELT). “An event in GDELT,” explains John Beieler, PhD student at the Department of Political Science at Pennsylvania State University, “is who did what to whom. Something like ‘Syrian rebels attacked the Syrian government’.” These events are logged “without human involvement,” says Beieler. Incidents all around the world are distilled from online news reports and then coded by a programme. The result is a massive database that reaches back to the year 1979 and is updated daily.
Using GDELT would therefore make it possible to do a rolling forecast and Ulfelder is currently experimenting with the data, but as Beieler points out there is at least one obvious problem with relying on media coverage. “The places we are most interested in − like CAR, South Sudan, North Korea − are the places that have the worst media coverage.”In some cases, this can make reliable forecasting impossible. “Could we have forecasted [based on GDELT] that South Sudan was going to happen?” asks Beieler. “I started looking at the GDELT data and it turns out that the answer is most likely ‘no’, because for many of the variables we look at, we had nothing.”
Another challenge of many data sources such as GDELT is how specific they can get. While in some cases such as Afghanistan, forecasting on the district level has been attempted, conflict forecasting for the moment is mostly limited to the national level as this is where most data is available.
With more input, however, Beieler argues that tools such as GDELT could help address these challenges. “[While] you can’t use GDELT really to pull out individual bits,” he says, “you can use it to create a thermometer of different places” such as whether a certain region is more violent this month than the last.
Now look here
With more granular and timely data, forecasts can only get more accurate. But as Ulfelder himself admits, the actual frequency of coups is very low and even those countries deemed to be at high risk are unlikely to experience a coup in reality. However, as Ulfelder also explains, the point of forecasting isn’t to get things right so much as to direct attention.
Rankings, by this argument, are a handy starting point rather than a definitive end in of themselves − they are a signal to international organisations, diplomats and NGOs that they should pay particular attention to certain places and, crucially, follow up with their own qualitative analysis. This is perhaps especially the case when the inclusion of certain countries defy conventional wisdom − in this year’s list, such examples might be Rwanda, Angola and Burkina Faso.
In this sense, statistical forecasting can serve to reduce blind spots and challenge unfounded narratives such as that of Mali’s democratic credentials before its coup. Once aware of conflict risks, the international community can pay closer attention, investigate further and take pre-emptive measures. Furthermore, even if the outbreak of conflict cannot be averted, that extra level of preparedness could prove invaluable in responding speedily and limiting the destruction, loss of life and fallout from the onset of a coup.