Will we see machine learning algorithms setting bank credit risk capital requirements?

Date23rd February 2022

Under Basel capital rules, accredited banks (‘IRB accredited banks’) can use internal models to determine their regulatory capital requirements.

Internal models used for calculating capital requirements need to be extensively validated, subject to review and challenge, as well as, understood by senior management. These models also need to be reviewed and approved by banking regulators. Internal models used for credit risk capital (‘IRB credit models’) use traditional model types such as generalised linear models (particularly logistic regression models) to predict default given borrower, collateral and loan characteristics. These model types are well understood, explainable and robust.

A tantalising question for banks and credit modelling teams is whether new machine learning techniques can replace traditional generalised linear models in IRB credit risk capital calculation. Banks are adopting machine learning in nearly all other aspects of their operations such as marketing, credit underwriting, pricing, and credit operations. When used with large volumes of data, these new model types can achieve higher accuracy relative to traditional models.

Some IRB accredited banks have experimented with machine learning techniques alongside logistic regression models or other generalised linear models. Machine learning models can be used as challenger models to benchmark IRB credit models. Alternatively, machine learning techniques can be used to identify combinations of features or even generate features which can be then used with generalised linear models.

Banking regulators have no doubt observed the growing use of machine learning within the banking industry and have been considering whether and how these models can be used in setting capital requirements. A recent consultation from the European Banking Authority (EBA) is, in my opinion, a step forward as the EBA formally seeks to understand whether machine learning models can be used for regulatory capital calculations.

EBA’s Discussion Paper and Consultation Process

The EBA is looking for responses from banks as well as banking supervisors on a range of questions related to applying machine learning models for IRB capital calculations. EBA accepted submissions until the 12th of February 2022. The consultation paper outlines machine learning models or techniques that can be used, some of the challenges that the EBA sees in applying these models in an IRB capital calculation context, an outline of some principles that it would expect to be applied should banks start incorporating machine learning models, and specific questions where it seeks feedback.

The EBA’s 17 questions in the discussion paper provide some insight on what the EBA is seeking to understand. These include:

  • Whether banks are considering using machine learning techniques for IRB models

  • What experience banks have in using machine learning techniques

  • What challenges do banks foresee when using machine learning techniques in areas such as data validation, interpretation, and implementation in IT systems

  • What are internal user challenges

  • How banks intend to share knowledge of these systems with senior management and use of explainability tools.

Within the discussion paper there are also some of EBA’s expectations. Obviously, any model used for IRB capital calculation will need to meet the requirements of its capital standards (EBA’s Capital Requirements Regulation). Further the EBA expects banks will also follow some of the principles for good implementation of big data and advanced analytics and it lists “data management, technology features, organization and analytics methodology”. EBA also calls out trust elements for machine learning and AI (ethics, explainability and interpretability, traceability and auditability, fairness and bias prevention/detection, data protection and quality, and consumer protection aspects and security).

A key theme reiterated in the discussion paper is that banks must be able to articulate the economic rationale for the model’s outcomes and predictor variables the model uses. Banks should also be able to apply human judgement in the model development process. This is challenging with machine learning techniques, where for example, interpreting hyperparameters is not straight forward. Finally, the discussion paper stresses that the models should be understood by senior management and users.

Conclusion

The EBA discussion paper is an important step in acknowledging the opportunity banks have to leverage machine learning techniques for IRB credit models. These techniques were not used in banks at the time the IRB framework was first implemented, however have now exploded in use in a range of areas within banks, including in credit risk measurement and management. We are currently working with consumer finance providers to develop machine learning models for credit underwriting and pricing. The models allow finance providers to use large datasets with a diverse range of predictors for default and customer behaviour such as early termination, and set prices at an individual borrower level.

The EBA discussion paper also outlines areas that should be considered more broadly when developing or adapting model risk frameworks for machine learning models. Machine learning models for credit underwriting in particular should have a rigorous control environment given the potential risk for lenders of getting these models wrong.

It is unclear whether banks in Australia would seek to implement machine learning models for IRB capital calculation in the short term. This may be seen as a ‘nice to have’ particularly when there are significant implementation challenges with perhaps limited tangible benefits such as lower capital requirements. Banks are handling other priorities such as the evolving COVID-19 pandemic, and regulatory change (including bedding down new capital standards).

Regardless of whether or not Australian banks will use machine learning techniques for IRB models in the near future, machine learning models and techniques will continue to be used for managing credit risk. IRB modelling teams should be familiar with the models and can potentially use these models to support the IRB model development process such as using these as challenger models or for identifying or generating features.

To read more about our Banking and Finance team, click here.

References

[1] Internal Rating Based (IRB) approach allows banks to use their own estimates for probability of default, loss given default and exposure of default for calculating regulatory credit risk capital requirements

[2] The terms “Machine Learning” or ML and “Artificial Intelligence” or AI tend to be used interchangeably. Machine learning technique and models include neural networks, gradient boosting machines and random forests.

[3] See https://www.eba.europa.eu/eba-consults-machine-learning-internal-ratings-based-models