You can’t miss the headlines in every newspaper at the moment, about the dangers that the world is facing from conflict: depending on which way you swing a lens, there is conflict in Ukraine, the Middle East, the Red Sea, South China Sea, and numerous places in Africa. Indeed, some have suggested that the globe is on the brink of World War Three. However, if you know just where the next conflict will break out, you have a much greater chance of stopping it, and that is exactly what Christopher Rauh is hoping to do.

Appropriately enough it is a cold dark stormy afternoon when I meet Chris at his office on the lower ground floor of the Faculty of Economics. I’m a few minutes early, and a passing faculty member comments that Chris will always be punctual – to the minute. And he is, having just arrived from a seminar.

He is part of a team that has investigated the signals that indicate a region might be about to turn ‘hot’, and result in civil commotion turning into either an all-out war, or at least a dispute that would result in fatalities.

Forecasting when an armed conflict is likely to break out is notoriously difficult, and history has shown conflict might suddenly appear out of a previously long lived and low-level dispute. “A special feature of our model is our reliance on newspaper text,” he says. “This allows us to capture changes in low level risk of conflict in countries which do not have a recent history of violence.”

I ask him to explain how and why he started researching the very specialised area of conflict.

“I was interested in because conflict is terrible, it destroys lives. It affects economic growth. It drives mass migration. So, it's important. And I was in a unique position where I could see the methodology. That came about because I worked in a consultancy where my job was to predict how many newspapers to put in each point of sale. And then when I started my PhD, I was surprised by how different academics approach these sort of problems,” he explains. “Often, they're much more interested in causal identification. But even when they were looking at correlations or descriptions, I had the fear they were not disciplining their models in the way you might have to if you're doing prediction. Quite simply I was interested in it and thought I might have a comparative advantage that I could share.”

However, I suggest that even the word conflict might be inflammatory. How do you actually define a conflict? I suggest a minor skirmish is very different to suddenly an all-out war breaking out.

“That's a very good question – and a difficult one too. There are a lot of models that have a different ‘threshold’ as we say. Some models are where there are more than 25 deaths in a battle, or 50, or more than a hundred. However much more useful is to use the same number of deaths per capita. Obviously, 100 battle deaths in, say, Luxembourg is different to 100 battle deaths in China. So, I take population divided by deaths.”

Then he explains how to predict the likelihood of it. “I particularly like predicting whether there is going to be at least one battle death because then there's no discussion about whether an event happened,” says Chris, saying that is his favourite model. “The other thresholds have fallen a bit out of fashion because of the arbitrariness. Then you come onto predicting the intensity, which is harder.”

However, I wonder, does this mean that the forecast of an intensity of a conflict is dependent on the language used by the newspaper. After all, saying there is likely to be a ‘civil war’ in Belgium could just be talking about divisions in the Parliament there. He explains that it needs to be coded quite carefully in the computer model.

“You're not going to get super red flags in any place just because some of those paper articles are talking about a sports event being extremely brutal, or a violent disagreement in a parliament.”

For example, each time they see a report about a war we look of the context of the text around it. “But actually, Belgium does have some cases that are coded in our data as war because of terrorist attacks. It's a question of definition. When Hannes Mueller, who is my co-author in these projects, and I launched the conflict website, on the first day there was surprise that Austria was dark red, which appeared strange – is Austria likely to be torn apart by war? I understand it looks strange, but that's because terrorist attacks are considered [as being] acted by a group against the government. And they're also very likely to happen again if they happen,” he says, showing me his screen, indicating that there was a high prediction of violence in Vienna and around Austria. “Equally, with Belgium, if you don't have any past violence, you're probably not going to have a predicted high value for a future conflict, so we code that in too.”

The newspaper texts that he uses work at the margin of these areas, and it also helps in locations where there is no indicator of past violence.

However, with terrorist attacks, I ask if there is a difference between lone wolf terrorist attacks and major conflict brewing – and how do you differentiate?

“If you gave me a data set and you remove all terrorist attacks, all I care about is a different type of conflict than terrorism. We look at whether there are rebel groups, for example. And our model trains on that very well, and it wouldn’t pick up on the terrorist case,” he says. “However, I still think it’s important, to train it our model including some of them because lone wolf attacks usually involve some sort of network, some sort of underlying sentiment which is shared by more people than just the lone wolf. It’s some sentiment that the Lone Wolf shares with some other people in that society. Therefore, you usually have more boiling under the surface. And the lone wolf is probably just the tip of the iceberg. We have to include them.”

He has produced a database of over 6 million news articles for gendered words like woman, man, boy, girl, and so on. This led to a data set of a monthly ‘maleness’ of country news coverage. This maleness is positively correlated with several topics such as ‘religious tensions’, ‘politics’, ‘foreign policy’, ‘armed conflict’ and ‘power and negotiation’ and negatively correlated with the topics ‘civilian life’, ‘health and emergencies’ and ‘sports’. He has not, however, managed to improve the conflict risk forecast through the gender index. “We find that the gender index does not change in a robust way before the outbreak of conflict either,” adds Professor Rauh.

I ask then, how do they filter out the low quality so-called ‘news’ website that in many cases are used to spread disinformation, often originating from Russia, compared with high quality newspapers that have a reporter on the scene.

“We have, and generally only use, vetted sources,” he explains. “The New York Times, The Economist, the BBC, Associated Press. These vetted sources have been around for a long time. Although of course these are Western sources, and we have to balance that to ensure we don’t have any bias. We supplement these because we don't have much Latin American coverage, and added Latin News, the news aggregator which only takes high quality news, from people who are there.

We just don’t use online blogs or news websites that have only been around for a few weeks or months, to ensure the sources are well vetted and consistent over time.”

Perhaps I suggest, if he is looking for more sources he might consider the text of broadcast radio or TV?

“Yes, we are looking at that,” replies Chris. “When we started this project, it was very difficult to convert speech into text and analyse it, but that is quite easy by now – it’s amazing how fast the IT world has moved on in just a few years. These things were nearly impossible a decade back. We are going to go there. But again, we’ll need a news archive going back to the 1990s which is searchable and downloadable. That costs a lot of money – and it’s debatable if we’ll get much more ‘bang for buck’ so to speak. But we'll get there.”

His new model for forecasting conflict has been partly funded by the Foreign, Commonwealth & Development Office (FCDO), emphasising the importance of the research on a strategic levels, for deciding on country priorities, and on the operational level, such as identifying critical periods by the country experts.

The real test for these models is to look back at historical conflicts. I ask how successful he has been feeding into the model with text from the past, and then seeing if conflict breaks out?

“It works very well,” he says. “That’s how we evaluate our models, in fact. For example, we pretend we are in January 2010, use all the information until then, and pretend like we don't know anything about the future, and predict it. On average, we do very well. But of course, most of this is driven by predicting violence where violence already happened.”

However, what is hard is predicting the sort of ‘Black Swan’ events in previously peaceful countries, when everyone is caught off guard. “It would be very interesting to go back to, say, 1913 and see if an incident in Sarajevo could have been predicted to have caused the July Crisis, and then the chaos of 1914 onwards. Alas we can’t do that yet with our data – it goes too far back. Although it could be feasible.”

However he cautions that these models find it hard to predict these sort of extreme escalations, because they are so rare. “We've ‘only’ had two World wars, so you can't build a predictive model on two occasions, alas” says Chris.
He did look at brute force pattern recognition for these one-off wars, but machine learning requires economic modelling where you're predicting something where you can’t train on past patterns.

“However, because of their huge significance, we have done some modelling combined with machine learning to be a bit more aggressive on predicting escalations because that's something that these models are very conservative on,” adds Chris.

The project has now developed to provide a complete ranking of over 170 countries with 65,000 grid cells across the world according to the risk that they experience an outbreak of armed conflict. A dynamic model which allows policy makers to trade off the many dimensions of costs caused by conflict with the costs of intervention under uncertainty.

I ask if there is an issue with predicting a negative – such as a conflict which was avoided at the last possible moment, such as the Cuban Missile Crisis.

“In a way we’re actually trying to use that for our advantage in the sense that we look at places with equal risk predictions in the past. And see where did it escalate and where did it not? And much more crucially, why. We use this approach to study the effect of power sharing agreements,” he says “So, we match countries on equal risk where they are both basket cases already. What you don't want to do is compare basket cases to functioning societies, and hence we match on equal risk and problems. And then, we can see if a power sharing agreements reduces risks. Like all these cases where we were under predicting conflict, there is clearly something to learn for policy, in particular whether diplomatic efforts work or where certain patterns helped peace, instead of war, to break out.”

The model which predicts outbreaks of violence and subsequent escalations into armed conflict, The Dynamic Early Warning and Action Model, by Hannes Mueller (IAE (CSIC) and Barcelona School of Economics), Christopher Rauh (University of Cambridge), Ben Seimon (Fundacio d’Economia Analitica) and Alessandro Ruggieri (CUNEF Universidad) simulates the costs and benefits of interventions. This should provide a testing ground for internal Foreign, Commonwealth & Development Office debates on both strategic levels, such as the process of deciding on country priorities, and on the operational level, such as identifying critical periods by the country experts.

“This allows the FCDO to simulate policy interventions and changes in its strategic focus. We show, for example, that the FCDO should remain engaged in recently stabilised countries and re-think its development focus in countries with the highest risks.,” he says. “The total expected economic benefit of reinforced preventive efforts, as defined in this report, would bring monthly savings in expected costs of 26 billion USD with a monthly gain to the UK of £520million GBP.”

I ask about recent events where we have gone in the past couple of years to not a lot of conflict around the world to suddenly everything happening from Ukraine, issues in Israel, the Red Sea, and a whole host of other flashpoints in many other parts of the world. Looking at his model, I can see that some areas are flashing bright red at the moment, so I ask Chris about the current state of play with the forecasts?

“Previous years might have been perceived to be calm, but that just shows that perhaps we have some perceived bias, given that it was just the Western world was calmer,” he replies. “There were still a lot of conflicts and battle deaths in sub-Saharan Africa and in some other places. They were flashing red well before then. And even within Ukraine, the Donbas and Lukhansk region were seeing small scale battles long before, along with the Middle East and in Israel. Sure, the scale in these two places is not particularly large, but they're not coming out of nowhere.”

Then we turn to the economic consequences of a conflict we're seeing in some parts of the world. In Ukraine for example every piece of infrastructure in some areas has been devastated. What can we say about the economic consequences?

He replies that it is something they’re working on much more now – but of course there are limits, as you can’t really use randomised controlled trials. “We cannot randomly assign conflicts to regions and countries to see what happens,” he acknowledges. “Instead, we rely on correlations and structural modelling, which clearly shows that conflict is in effect a gigantic tax. It’s much worse than for example homicides, which are not reflected in huge GDP growth losses. If you look at, say, Brazil, it has had more people dying than in Afghanistan. It's terrible. But it doesn't have the same devastating effects to growth as battle deaths. Having even a small war is like having a huge haircut to your growth.”

“We're also trying to study if you're in these intermediate stages, what growth looks like. And we see those places with high risk predictions are also seeing growth losses, coming out of future potential conflict,” says Chris. “That reinforces our point. You really want to avoid that first outbreak because there's such a huge cumulative economic loss in the future.”

As storm clouds gather outside the Faculty windows, we turn to how he will develop the research.

“There are two fronts. I feel it is crucial to explore the role of policy in reducing risk. That is actually harder than you’d imagine, as there are no good databases on it. Saying for example this country did one thing, and another took a different tack. We talk a lot with foreign offices that just don't have very good systematic data of their efforts.”

“Then I’d like to go a bit crazy on the prediction side of leveraging all the new developments in artificial intelligence. Different ways of working with text, different ways of predicting, there is a lot more that can be done.”

Scary though it sounds, it does sound like there is a lot more work ahead, and as the world hangs on the edge of another big conflict, it sounds like developing his model will keep Professor Rauh busy.

And as I get up off the sofa in his office, appropriately enough, the clouds outside the window seem to darken ever so slightly.

Professor Rauh’s paper on ‘Building Bridges to Peace: A Quantitative Evaluation of Power-Sharing Agreements’ was published in January: https://academic.oup.com/economicpolicy/advance-article/doi/10.1093/epolic/eiae010/7587518