Answering the Queen: Machine learning and financial crises

BIS Working Papers  |  No 926  | 
01 February 2021



Financial crises wreak economic, social and political havoc. They impose large fiscal costs, driving up public debt and disrupting our societies. Designing and implementing policies that reduce the chances of a crisis occurring – as well as its costs – requires the build-up of risk in the economy to be recognised at an early stage. This paper is an attempt to restore some imagination to the economics of crises.


We adapt the framework of sequential prediction or online machine learning to forecast systemic financial crises without knowing the "true" model of the economy. Our approach can be described as "meta-statistic", since the aim is to make the best possible prediction by aggregating predictions from different models. These models are estimated using the standard macroeconomic variables used in past studies. Our methodology and results may be useful for the conduct of macroprudential policy.


We uncover a time-varying subset of models that carry most of the information needed to predict financial crises. Among those models, we also discuss which ones "flash red" at the right time. Using a mix of 26 models for France, Germany, Italy and the United Kingdom – including central bank financial crises models as well as machine learning models – we are able to predict a systemic financial crisis three years ahead out-of-sample, with lower signal-to-noise ratios than in the existing literature.


Financial crises cause economic, social and political havoc. Macroprudential policies are gaining traction but are still severely under-researched compared to monetary policy and fiscal policy. We use the general framework of sequential predictions also called online machine learning to forecast crises out-of-sample. Our methodology is based on model averaging and is "meta-statistic" since we can incorporate any predictive model of crises in our set of experts and test its ability to add information. We are able to predict systemic financial crises twelve quarters ahead out-of-sample with high signal-to-noise ratio in most cases. We analyse which experts provide the most information for our predictions at each point in time and for each country, allowing us to gain some insights into economic mechanisms underlying the building of risk in economies.

JEL codes: E37, E44, G01