I have a new paper with Christopher Gandrud on prediction and financial crises. The genesis of the paper lies in the debate about why the Global Financial Crisis caught so many people, especially social scientists, by surprise. One argument is that social scientists are stupid and/or naive, or worse, that they were just cheerleaders for finance. Another argument is that crises are just unpredictable, period; indeed, if the efficient market hypothesis is valid then price movements, even large ones, should be unknowable ex ante.

Without denying either of those possibilities (speaking for myself at least), we are interested in a different argument, which is that financial crises are unpredictable because they are subject to self-fulfilling dynamics. Our paper shows how to model a situation in which market actors play a coordination game with private information about their “strength” or risk tolerance. Crises emerge when players both sell an asset, and both strictly prefer to hold the asset when their counterpart holds too. We ask, given that there is not a crisis in the present, what can we infer about whether a crisis will happen in the future? The model generates a novel prediction: that the better economic conditions are in the present, the less we know about the likelihood of financial crises in the future. Looking at a new dataset that Christopher and Mark Hallerberg collected, we find that the data bear out this prediction.

We see our argument as complementing Timur Kuran‘s work on “preference falsification,” which also shows how a simple model of social interactions explains why it should be hard to predict things like revolutions, uprisings, and crises. It also allows us to relate prediction problems to “epistemic uncertainty” (in a Knightian sense) in a single framework. There are also implications for what regulators ought to do if they want to minimize the threat of crises: learn more about market participants’ ability to withstand negative shocks. Stress tests, for example, are one way to do this.