IRB models using advanced Machine Learning methods, EBA opens the door with their new discussion paper. Read about entering this next era in IRB modelling.
EBA recently published a discussion paper on Machine Learning (ML) for Internal Ratings-based (IRB) Models (EBA/DP/2021/04). The aim of the paper is to set the supervisory expectations on more sophisticated ML models and their adherence to the Capital Requirements Regulation (CRR) in the IRB models context for risk management. In the paper, machine learning is scoped to be more advanced methods. For example, tree methods, neural networks (NN), generalised regression and boosting methods are mentioned.
Up until recently, regulators have been wary of the use of more complicated ML models (see ‘EBA Report on Big Data and Advanced Analytics‘ and IIF’s ‘Machine Learning in Credit Risk‘). Potential issues related to explainability, ‘black box’ approach and limited understanding of theoretical frameworks are highlighted. The discussion paper touches upon all these points.
It is great to see regulators are opening to the idea of using ML, as such methods are becoming easier to use and implement. However, there are still a lot of challenges to tackle in the use. Most advanced ML methods require more high-quality data. Besides, these methods require careful calibration of the models to get the best performance, without overfitting the data.
Additional challenges remain in terms of:
- Organisations’ readiness for more advanced models
- High level understanding of the methodology by the senior management
- Interpreting the results and justifying the use within the organisation and to the regulators
ML Methods Already Bringing Benefits to Model Development
EBA identifies several potential benefits from the use of Machine Learning (ML) models, ranging from data preparation, model development (e.g., risk differentiation and risk quantification), validation and monitoring of models. Such methods are already being used or being tested in financial institutions. Potentially, they will play a bigger role now the regulator is opening up to the opportunities of ML. Right now, we already see developers testing their usual methods with more advanced algorithms to detect outliers. For identifying groups of clients with similar trends they use clustering algorithms to gain insights into client behaviour. Financial institutions also use advanced ML models for internal model submissions. For example, model validation (MV) departments request or build challenger models themselves using ML methods.
We also have discussed potential use cases, particularly in a LGD model, in our previous insights (part 1, part 2). The analysis showed that the ML algorithms are data hungry and require detailed investigations and understanding to justify the use in a credit risk model.
Next to the numerous possible benefits from the use of ML methods, internal models used for capital purposes of course require supervisory approval. Hence, the models need to comply with the regulatory requirements, such as the CRR. EBA therefore provides a list of principles, as a guidance to institutions in the context of assessing whether ML models can be ultimately approved by the supervisors.
Going to the Next era in IRB Models, Using Advanced ML Methods
EBA notes the four pillars for the development, implementation and adoption of big data and advanced analytics in the form of data management, technological infrastructure, organisation & governance and analytics methodology. These are all necessary to support the rollout of advanced analytics. However, the EBA also puts emphasis on several trust elements. For example, ethics, explainability and fairness shall be sufficiently addressed when implementing Machine Learning (ML) models. In an earlier insight, we describe how to add fairness to Data Quality Frameworks. This is becoming more and more important as governing bodies are focusing on AI and ML methods used in decision making.
To conclude, onboarding sophisticated ML methods for regulatory capital calculation in your organisation needs experience and vision in different topics. At Amsterdam Data Collective, we have extensive experience in credit risk, IRB model development and model validation. We have also been using advanced ML models on various business use cases across different sectors, such as the healthcare and public sector. Amsterdam Data Collective can help you assess readiness for such ML models. We are also already using different ML techniques in credit risk context for data quality, variable selection, prototyping, model monitoring and as challenger models. We have also touched upon these points in our recent blog post.
How can Amsterdam Data Collective help?
Would you like to know more about the role ML Models can play within the area of capital models? Or what the ADC can do for you? Please contact Scott Bush at firstname.lastname@example.org or check our contact page.