March 2022

The use of Machine Learning (ML) in Risk Management and, in particular, in IRB systems has sparked a heated international debate, as evidenced by the various publications on the subject from both the world of academia and the banking market. In terms of banking, the European Banking Authority published a consultation paper on the possible use of ML for internal ratings-based (IRB) models in November 2021.

At the heart of the discussions are some fundamental issues concerning the adoption of machine learning methodologies:

  • potential benefits, i.e., the modeling contexts in which they create value, such as the potential offered to data enrichment by unstructured information and the supporting role of increasingly digital processes;
  • direct and indirect costs, such as those linked to the acquisition of new analytical skills or new information assets or the costs of inaction, or the failure to evolve toward such methodologies in an increasingly competitive and regulated market;
  • compliance with regulatory requirements (e.g., the IRB model framework in prudential regulation, GDPR, etc.);
  • the level of confidence in such methodologies with particular regard to the robustness of model performance (management of overfitting), and the ability to “understand” the model (the so-called interpretability) and to govern all the steps of the development process (human oversight of the process) to be able to prevent ethics and fairness issues related to the use of the ML model.

With the aim of providing greater clarity and to respond to the challenges posed by the adoption of Machine Learning in Risk Management, Banca Intesa Sanpaolo and CRIF have shared their experience in a position paper entitled “Machine Learning for Credit Risk management and IRB models: lessons from successful case histories”.
After a brief introduction to machine learning and the potential relationship with the banking market, the paper proposes a careful analysis of the reference regulatory context to respond to the “practical” challenges of adopting such methodologies through five case histories.
The work carried out by CRIF-Intesa Sanpaolo was presented to the European Banking Authority (EBA) at the end of March 2022.