The European Central Bank has published a working paper developing a granular machine-learning framework that estimates European non-financial corporations’ probability of default and transmits stressed firm-level risk to banks through loan exposures. The analysis finds that random forests outperform logistic regression in predicting defaults and are more sensitive to adverse scenarios, capturing nonlinear and heterogeneous effects as well as banks with elevated tail risk. Using financial data for a representative sample of 10,000 firms, scenarios from the 2023 European Union-wide stress test and AnaCredit exposures, the framework maps macroeconomic and sectoral shocks into firms’ financial statements before calculating bank-level risk measures. Under the adverse scenario, the aggregate probability of default roughly doubles in the first year to about 7%, while the random forest-based bank riskiness index rises from 5.5% to more than 8.5% during the first two years. Default risk responds particularly strongly when interest rate spread shocks cross certain thresholds, while gross value added shocks produce a weaker nonlinear effect.
2026-08-24European Central Bank
European Central Bank working paper finds random forests improve corporate default and bank stress testing
A European Central Bank working paper finds that random forests outperform logistic regression in predicting and stress testing European corporate defaults. The framework links firm-level stress to banks through loan exposures and captures nonlinear scenario effects and tail-bank risks, with aggregate default probability reaching about 7% in the first year of the adverse scenario.