Credit decisions with artificial intelligence and model risk

What it is about

How artificial intelligence is transforming credit risk management through models and advanced analytics that strengthen decision-making, and the key challenges and best practices in model risk for a responsible and reliable use of these technologies.

Session summary

The first part reviews how machine learning applies across the credit life cycle: origination, monitoring, early-delinquency and recovery models, all built to estimate a probability tied to a target, such as default within the next twelve months. It recalls the basic expected-loss formula (probability of default, loss given default and exposure) and summarizes a literature review according to which machine learning models predict better than traditional logistic regression, with tree-based models doing better than neural networks. It also warns about the challenges: model stability, overfitting, data governance, bias, privacy and the difficulty of explaining ‘black box’ models, for which it mentions post-hoc interpretation techniques such as SHAP.

The second part defines model risk as a cross-cutting risk with three sources: the data, the methodology and the use made of the model. It proposes a governance framework with a model inventory approved by the board, independent validation, documentation, roles for the three lines of defense, classification of models into risk tiers and performance monitoring throughout the life cycle. It closes with the risks that artificial intelligence adds and five conclusions: managing model risk is a condition for scaling its use, and data are the main determinant of quality. The questions covered bias, prohibited variables, model committees, the profile of those who develop and validate models, and including model risk in the risk appetite and in the capital self-assessment.

Topics covered

• Machine learning models across the credit life cycle
• Expected loss and probability of default: comparison with logistic regression
• Explainability of black-box models and bias
• Model risk: sources, life cycle and governance
• Model inventory, risk tiering and independent validation
• Risks added by artificial intelligence and data quality
• Questions: prohibited variables, model committees, risk appetite and capital