BoE Paper on Machine Learning Approach to Predict Financial Crisis

The paper describes the dataset and reviews literature on the variables that the authors choose as predictors. It then presents the benchmark logistic regression, outlines the methodology, and provides a brief description of the different machine learning models applied and the Shapley value framework. Next, the paper compares the predictive performance of all models and investigates the importance of the predictors using Shapley values. Finally, the paper looks at the robustness of indicators in a macro-prudential policy context and investigates in detail the role of the yield curve. 

This paper shows that machine learning models outperform logistic regression in predicting financial crises on a macroeconomic data set covering 17 countries between 1870 and 2016 in both out-of-sample cross-validation and recursive forecasting. The gains in predictive accuracy justify the use of initially more opaque machine learning models. All models consistently identify similar predictors for financial crises, although there are some variations across time reflecting changes in the nature of the global monetary and financial system. While the crucial role of credit is an established result in the literature, the predictive power of the yield curve has obtained far less attention as an early warning indicator. The authors also inspect non-linearities and interactions identified by the machine learning models. Global credit shows a particularly strong non-linearity—only very high global credit growth beyond a certain point influences the prediction of the models. Interactions are particularly strong between global and domestic indicators.

Overall, the findings suggest a combination of low risk perception, search-for-yield behavior, and strong credit growth in the years preceding a crisis. The results highlight the potential value of machine learning models for broader economic policy making in two key dimensions. First, the approach illustrates how machine learning techniques can uncover important non-linearities and interactions that facilitate superior out-of-sample prediction and forecasting even in situations characterized by relatively small data sets with limited observations of the event of interest. Second, the novel Shapley value approach demonstrates how the black box concern linked to the practical policy application of machine learning models may be overcome. By providing a mechanism to identify the key economic drivers of the predictions generated by such models, the approach allows insights from machine learning models to be integrated into a broader decision making framework while preserving the transparency and accountability of any resulting public policy decision.

Related Link: Working Paper

Keywords: Europe, UK, Banking, Machine Learning, Yield Curve, Financial Stability, Dataset, Research, Early Warning Model, Fintech, BoE