Competition Highlights Work of Georgia Tech Researcher Helping Drive Improvements in the Science of Conflict Forecasting

It’s not every day a social science scholar gets a chance to shape a field from nearly the ground up, but that’s the exciting reality in which David Muchlinski finds himself. 

The assistant professor in Georgia Tech’s Sam Nunn School of International Affairs is one of only about 40 people worldwide working in the young field of conflict forecasting, with hopes of one day advancing models that could give governments, international organizations, NGOs, and others the kind of insight to help them head off conflict, not merely respond to it. 

It’s an exhilarating high-wire act for a young researcher living through a tumultuous period in world history. 

“Knowing I could make a difference in some way in the world is cool to think about, but this is also a scientifically difficult and fascinating problem to crack,” he said. “You’re forced to put actual predictions out there and be judged by history about how right or wrong you are or were. Relatively few researchers in political science are willing to put their reputations or the validity of their models on the line in such a public way and be proven wrong, sometimes very wrong.” 

Conflict Competition Seeks to Drive Innovation 

That’s exactly what he’s doing now as part of a triennial academic competition to drive improvements in the science of conflict forecasting. 

Muchlinski and Ph.D. student Chandler Thornhill’s contribution to the triennial Violence and Early Warning System (VIEWS) Prediction Challenge involves a new two-step model that the researchers hope better addresses common challenges in violence prediction and outperforms existing models. 

"One of the primary challenges in conflict forecasting is that, fortunately, there are relatively few cases that researchers are able to study,” Thornhill said. “If we are looking at all states across multiple years, the number of states experiencing a conflict in a given year is relatively small, meaning a lot of zeroes in the data. This is where the importance of a two-step model comes in.”  

The first step in their model evaluates whether a given country is likely to experience conflict. The second step estimates the level of intensity of conflicts in countries where it is predicted to occur. 

Muchlinski and Thornhill say their model predicts Ethiopia, Ukraine, Yemen, Afghanistan, and Israel will lead the world in armed conflict fatalities during the competition, which runs through May. 

History of Conflict Prediction 

The field dates back to the 1980s, with scholars such as Bruce Bueno de Mesquita of New York University and Philip Schrodt, a senior research scientist at Parus Analytics and former professor at Penn State and the University of Kansas. 

“They led early efforts to collect numerical data on conflicts and wars, systematize this data into relational databases, and use computer and/or statistical modeling to predict future conflicts,” Muchlinski said. “These early efforts were often sponsored by agencies like the CIA, and much of the early models were proprietary and not subject to public scrutiny.” 

The current field began to emerge only in the 2010s as scholars, including Muchlinski, began using machine learning techniques to advance their predictive capabilities. It was during this time that Muchlinski published his first paper on the topic, “Comparing Random Forest with Logistic Regression for Predicting Class-Imbalanced Civil War Onset Data” in 2016, in Political Analysis. 

Machine learning techniques and large language models (LLMs) that power AI chatbots such as ChatGPT are rapidly taking hold in the field, promising greater predictive capabilities, Muchlinski said. But many challenges remain. 

“Wars are complex processes,” he said. “While we theorize that some factors may be important, like weak states generally don’t attack stronger states, or that states declining in power may launch preemptive wars before they become too weak, these explanations generally explain only a handful of cases, or are so commonsensical and outside the realm of good scientific theory as to be meaningless.” 

“We don’t know many things that may cause conflict, and we don’t know where to look for answers,” he said. “And even worse, we don’t know if explanations for one conflict explain others, or if each conflict is more or less unique and explained by idiosyncratic factors.” 

What’s Next for the Field? 

Muchlinski said conflict prevention scholars are driven not only by the challenge but also the possibility of someday arming governments and others with the information and time they need to try to stave off conflicts instead of simply responding to them. 

“The ultimate goal of our work is being able to effectively forecast with at least a year’s notice that something unforeseen like the current Israel/Hamas or the Russia/Ukraine war will occur,” Muchlinski said. 

The use of LLMs shows good promise in helping researchers better understand problems like these, Muchlinski said. But the field also needs more strenuous theory development and testing, which competitions like the one Muchlinski and Thornhill are working on may help drive. 

But even if they get it right and learn to make accurate, timely predictions, it’s still going to be on world leaders to solve problems before they get out of hand, Muchlinski noted. Data science can’t solve that problem, he said. 

“No statistical model will convince reluctant politicians to loosen the purse strings. Ultimately that’s where the difference would lie, at the intersection of forecasting and political will” he said. “One of those is a much easier problem to solve than the other.”