This paper studies how self-fulfilling dynamics affect the predictability of financial crises. We build a model in which market participants play an investment coordination game with common economic shocks and private information about their own willingness to cooperate. An observer attempts to predict the occurrence of a financial crisis in the future based on the players' actions in the present. Under favorable conditions, good market conditions prevent an observer from learning about players' types by observing their actions. This induces a negative correlation between economic conditions in the present and the variance of players' types in the following period, which defines the predictability of financial crises. Under more extreme positive market conditions, the observer cannot use present observations of the players to learn about the probability of a financial crisis in the future. We test the implications of the model using a new continuous measure of financial market stress - “FinStress” - developed by Gandrud and Hallerberg (2015). We find support for a key implication of our theory: the variance of financial market stress is larger following periods of good economic conditions than following poor economic conditions. These conclusions have implications for both empirical analyses of the predictors of financial crises and for policymakers seeking to prevent crises. The findings in our paper suggest that regulators should focus on actors' types (e.g. with stress tests) more than macro-economic conditions. More broadly, our analysis provides a new explanation for why social scientists and area specialists are generally poor at predicting events that require mass coordination, such as financial crises, coups, and revolutions.