Random forests reveal non-linear tail risks in bank stress tests
A machine learning framework using random forests improves corporate default predictions and reveals non-linear tail risks across European banks. Research published by the ECB shows the algorithm outperforms standard logistic models across 10,000 firms and 2,039 lenders.
Superior prediction across corporate balance sheets
Evaluating a representative sample of 10,000 euro area non-financial firms from Orbis, the random forest model outperforms logistic regression across all horizons.
The model achieves an average out-of-sample area under the precision-recall curve of 0.664 compared to 0.577 for logit, alongside a higher ROC-AUC of 0.967 versus 0.917. SHAP feature attributions show that cash-flow and liquidity variables—namely financial expenses, total cash, and operating cash flows—dominate default prediction over aggregate macro indicators.
Under the adverse scenario from the 2023 EU-wide stress test, projected default rates roughly double in the first year, rising by 2.5 percentage points to approximately 7 percent.
Mapping exposures across the banking network
Mapping corporate default probabilities into an AnaCredit bipartite network of 6,215 firms and 2,039 banks reveals critical transmission channels.
The dataset covers 88 percent of AnaCredit institutions and 73 percent of banking assets.
Random forest models capture non-linear tail risks where standard logit fails: when spread add-ons exceed the 75th percentile, default rates jump abruptly, identifying vulnerable tail lenders whose systemic riskiness exceeds 80 percent.
Sharp diagnostic, operational hurdle
The framework proves that standard linear models dangerously understate tail risk during financial stress.
Yet massive data requirements and frozen balance sheet assumptions restrict immediate supervisory deployment.
Broad adoption will hinge on whether regulators truly trust complex machine learning over simpler formulas.