“Because war is costly and risky, rational states should have incentives to locate negotiated settlements that all would prefer to the gamble of war.”James D. Fearon · Rationalist Explanations for War · 1995 · James D. Fearon, «Rationalist Explanations for War», International Organization, vol. 49, nr. 3, 1995, pp. 379–414 (formularea din corpul articolului, redată din memorie; traducerea noastră)
War comes not because the sides would not want a deal, but because they cannot believe one.
In «Rationalist Explanations for War» (International Organization, 1995), James Fearon started from an awkward question: war destroys resources, so both sides should prefer a negotiated settlement. If they fight nonetheless, something prevents a deal. Fearon identifies three rational explanations: private information with incentives to misrepresent it, commitment problems and the indivisibility of the stake. The other popular explanations (irrationality, mistakes, domestic politics) are not rejected, but these three can be modelled without assuming leaders do not think. The mechanism. Private information: each side knows its own capability and resolve better than the adversary does, and each has an interest in seeming stronger; so statements are not trustworthy, and fighting sometimes becomes the way of finding out. The commitment problem: a deal made today does not hold if one side will grow stronger tomorrow, so the other prefers to strike first. Indivisibility: some stakes (a holy city, a throne) cannot be shared, though Fearon notes that compensation can usually be negotiated. Limits: the model assumes unitary, rational actors, and for forecasting the question remains how you measure ‘private information’. What it means for the forecaster: wars are rare, and most disputes end without fighting, so the base rate is low; it rises when signals appear that remove uncertainty (costly mobilisation) or when a rapid shift of power makes a commitment no longer credible.
Worked example: two years of signals before 24 February 2022, and the arithmetic of a base rate. The Correlates of War (COW) list, version 4.0 for 1816–2007, counts 95 interstate wars, defined by at least 1,000 battle deaths. The calculation: the period 1816–2007 has 2007 − 1816 + 1 = 192 years; 95 divided by 192 gives about 0.49 wars a year, so on average one every just over two years across the whole international system. For a specific pair of states the annual rate is incomparably smaller: these 95 wars are spread across thousands of dyad-years.
Now the case. In spring 2021 Russia massed troops near Ukraine, then withdrew some of them in April, without fighting. The alarm proved false. In autumn 2021 the build-up resumed, larger, and lasted for months; on 24 February 2022 the invasion began. From Fearon’s perspective the difference between the two episodes is not the number of troops but the information they revealed. A cheap signal (a deployment that can be cancelled with no political loss) does not remove uncertainty; a costly one (months of visible preparation, maximalist public demands) does, because the sender loses much if he backs down.
What the case shows a forecaster: there is a low base rate (interstate war is rare), but costly signals update that rate strongly. After the 2021 build-up many analysts put small probabilities; after months of visible preparation a few raised them sharply. No single signal decided; it was the accumulation of those that cannot be withdrawn for free.
The lesson: ‘mobilisation at the border’ does not mean invasion, but it is not zero either. The right question is what the signal costs, not how large it is.
For the forecaster: a reference class for ‘the outbreak of an interstate war’ exists: Correlates of War (the ‘Inter-State Wars’ list, 1816–2007) with 95 wars and a threshold of 1,000 battle deaths, complemented by the UCDP/PRIO series, with a threshold of 25 deaths a year. The two thresholds give different base rates, so the question must name the source and the threshold.
Indicators, with the source: the UCDP database (Uppsala University) for active armed conflicts; defence ministry reports and commercial satellite imagery for deployments; official statements with a deadline and maximalist demands; the pace of arming (SIPRI). For each, you note whether the signal is cheap or costly.
The typical trap: confusing a dispute with a war. Most territorial disputes never led to fighting, and a base rate computed only on the cases that escalated is inflated and misleading. A second trap: taking every build-up as an invasion or as a bluff; both happen, and the difference lies in the cost.
Example scorable question: ‘Will the version published in 2028 of the UCDP/PRIO Armed Conflict dataset contain at least one interstate-type conflict active in calendar year 2027?’ It resolves yes/no from the dataset; if UCDP does not publish the version by 31 December 2028, the question is voided.
Why it matters Russia’s attack on Ukraine of 24 February 2022 followed a troop build-up in spring 2021 that ended without fighting, so the same signal was a false alarm once and correct the second time.