Machine learning in credit granting: an analysis of the opportunities

What it is about

Use of machine learning models for credit risk management, in particular the prediction of the probability of default, and for economic capital.

Session summary

The speaker, a financial innovation researcher at the Bank of Spain, first clarifies what is meant by artificial intelligence and machine learning: he distinguishes models based on assumptions about the data (traditional econometrics) from algorithm-based models that learn from the data, and recalls that European artificial intelligence regulation classifies lending as a high-risk use. He explains why credit is the natural ground for studying these techniques: plenty of data, non-linear behavior and a default variable that can be classified.

He presents a comparison of six models, from penalized logistic regression to deep neural networks, on a real consumer-credit portfolio. Tree-based models, in particular gradient boosting and random forests, discriminate and calibrate better, and he separates the improvement that comes from having more data from the one that comes from the model itself. When the default probabilities are fed into the Basel internal-models capital calculation, he estimates a capital saving of between 12.5% and 17% compared with the model most used by the industry, which he considers a lower bound. The questions covered other definitions of default, replicating the exercise outside Basel, the useful life of models, putting them into production, data governance and the use of unstructured data such as text.

Topics covered

• What machine learning is compared with traditional statistics
• Credit as a high-risk use in European artificial intelligence regulation
• Comparison of six probability-of-default models
• Data advantage versus model advantage
• Calibration and regulatory capital savings with internal models
• Scalability, data governance and putting models into production
• Questions: definition of default, useful life of models and unstructured data