Revenge of the Experts: Building Better Conflict Predictions

When a crisis breaks—troops amass along a border, a ceasefire wobbles, a leader issues an ultimatum—journalists often contact experts for comment, often academics who study geopolitical conflict. Increasingly, though, the question of whether a conflict will turn violent gets answered somewhere else first: on Polymarket, Kalshi, Metaculus, and similar platforms where users make bets (or as insiders say, ‘trade contracts’) on whether a ceasefire will hold, whether one state will strike another, or whether an invasion will begin. A contract priced at 30 cents means the crowd thinks the event has a 30 percent chance of happening.

The appeal of a prediction market is that those predictions will be more rigorous and carefully made, since users own money is at risk. A pundit can say war is “likely,” later deny having meant anything specific, and lose nothing. A forecast tied to money should force more discipline into the claim by raising the stakes—and in so doing, be a useful input for policymakers. But none of this is true if the forecast starts with bad information.

That baseline information is what social scientists call a base rate. Base rates explain how often a thing (like a war between two countries) usually happens before experts add the special facts of the case at hand. Doctors use base rates all the time. A patient who looks pale may be in trouble, but a good diagnosis begins by asking how often patients of a similar age, blood pressure, medical history, and symptoms actually suffer the feared event. Without that comparison, vivid evidence can lead experts to the wrong conclusions.

Conflict forecasting has the same problem. Troop movements, threats, missile tests, and diplomatic deadlines should all factor into a risk assessment. But they should be considered alongside probabilities based on historical evidence. If analysts leap straight from dramatic headlines to dramatic conclusions, they commit the base-rate fallacy: treating the most visible facts as if they were the whole evidence base.

However, when it comes to predicting war, the most obvious base rate—how often do any two countries fight in a given year?—is almost useless. The answer is almost never. There is no impending Indonesian invasion of Bolivia; no air war between Kyrgyzstan and Brunei. Most pairs of countries do not fight because most pairs of countries have no active dispute, no mobilized forces, and no urgent decision in front of them. Starting from all country pairs makes almost every warning look hysterical. This is especially worrying for academic forecasters since most of the published statistical models of conflict involve every country dyad on the planet, even the overwhelmingly mundane ones.

Real life analysts rarely ask, “Will a random pair of countries go to war?” They ask a conditional question: Now that troops are moving, leaders are issuing deadlines, sanctions are escalating, or shots have been fired, how often do situations like this get worse? The relevant comparison is not normal peacetime—which is what most academic statistical models are built on—but the subset of dyads already conditional on being in what is known as a “crisis”.

That is where international relations research helps. The International Crisis Behavior (ICB) project, which analyzes data on international crises and crisis actors since the end of World War I, and was built by political scientists Michael Brecher and Jonathan Wilkenfeld, among others, is a useful example. The project doesn’t predict wars with mechanical certainty; it defines a crisis in a careful way. A case enters the dataset when leaders perceive a serious threat, have limited time to respond, and when there is a real possibility of military hostilities.

The ICB definition of conflict induces a salutary selection effect by ignoring prosaic day-to-day behavior between states. The ICB data are not a catalog of every hostile speech, border incident, or provocation that fizzled out. They are a record of disputes that had already crossed a meaningful threshold. Once conditioned on that threshold, the world looks very different. So what are these crisis-conditional base rates every scholar of international relations ought to know? In the post-1945 system-level data, roughly a quarter of crises involved no violence, about a third involved minor clashes, another quarter involved serious clashes, and 17 percent reached full-scale war. Violence is not automatic once a crisis begins. But it is not a remote abstraction either.

Crises that begin with direct or indirect violent acts escalated to serious clashes or full-scale war about 61 percent of the time in the post-1945 ICB data. Crises triggered by nonviolent military moves, such as mobilizations or shows of force, escalated at about 19 percent. Crises that began with verbal, political, or economic acts escalated at about 23 percent. The opening move is therefore an important clue about the likely path of the dispute.

Perceived stakes matter, too. When leaders saw the threat as existential, crises escalated to serious clashes or full-scale war about 69 percent of the time. Lower-gravity crises did so about 37 percent of the time. In plain English: a crisis that feels existential to one side is different from one involving limited stakes, even if both produce alarming headlines. Yet it is also worth noting that such crises “only” increase the base probability of conflict to 70 percent, not 100, since states can still come to a negotiated settlement or drop the issue entirely.

Once a dispute has crossed into a crisis in that sense, the questions that matter aren’t a generic list of things to notice. They involve looking at the documented record of what each side actually did, not from guessing what anyone privately believes. What triggered the crisis, and what does that reveal about how resolved each side actually is, as opposed to how resolved they claim to be? What is actually at stake, and what is that worth relative to the cost of testing the other side’s commitment by force? Has any action been taken that a cheap announcement can’t fake—a troop movement, a stood-down asset, a broken deadline? And just as important: can a deal reached today actually hold? Deals don’t only break down because one side is bluffing about its resolve. They also break down because a concession made now can’t guarantee concessions and cooperation once power, alliances, or intentions shift later—which is why “we’ll settle this diplomatically” often isn’t enough on its own.

Better base rates will not make conflict (or peace) entirely predictable. They will not remove moral concerns about betting on violence, nor will they solve problems of manipulation, thin markets, or insider information. But they can make experts more relevant in a world in which these markets are ubiquitous. They can show journalists when a dramatic headline is mostly noise, help analysts separate background rivalry from acute danger, and give policymakers and humanitarian actors a clearer sense of where attention is needed before violence arrives. Policymakers still turn to experts when a crisis breaks. Sharper expertise means a better chance of that response arriving in time to matter.

Adi Rao is a postdoctoral fellow in technology and international security at the UC Institute on Global Conflict and Cooperation (IGCC) in Washington, D.C.

Thumbail credit: Unsplash