Use of AI in risk management
The Artificial Intelligence Round Table aims to discuss developments in artificial intelligence (AI) and its opportunities and challenges for risk management with banking practitioners and AI experts. In view of the increasing importance of AI for banks and financial service providers, it is essential to understand the opportunities and risks of this technology and to develop suitable risk management strategies. The round table organized two discussion rounds and produced a position paper on the topic of artificial intelligence in model risk management.
Presentation of the LGD Challenger model
In the meeting on February 29, 2024, the Advisense team presented the AI-based LGD benchmark model. The so-called Challenger model uses machine learning techniques such as XGBoost to improve the accuracy of loss-given-default (LGD) estimates. The model was developed with over 50 risk drivers and enabled a more precise prediction by segmenting the LGD cases into different categories.
The tests showed that significant improvements in model development could be achieved through the use of AI. These results illustrate the potential of AI to identify new risk drivers that cannot be recognized using conventional methods.
Another focus was the management of risks arising from the use of AI and ML models. Matthias Fahrenwaldt from BaFin highlighted the challenges posed by the complexity of the models and the deviation of real market conditions from model assumptions. The discussion made it clear that an expansion of the existing model governance framework and the development of new validation approaches are required.
Use of unstructured data
In June 2024, Vahe Andonians and Ara Abrahamyan from Cognaize presented their approach to leveraging unstructured data by combining traditional AI models with generative LLMs. This approach enables more efficient use of information from various sources such as annual reports and news feeds and offers significant advantages in AI-supported risk management.
A central aspect of the sessions was the question of how AI can change risk management as a transformative meta-technology. Philipp Adamidis from QuantPI emphasized that the centralization of AI governance and testing as well as the automation of validation processes are crucial to cope with the growing requirements.
Position paper published
In September 2024, the Round Table published its position paper on artificial intelligence. The paper was written by Dr. Sebastian Fritz-Morgenthal (Advisense), Philipp Adamidis (QuantPI) and Dr. Jochen Papenbrock (NVIDIA) and provides a comprehensive analysis of the regulatory requirements as well as the challenges and opportunities of using AI in model risk management.
The position paper emphasizes the strict requirements imposed by the EBA, ECB, BaFin and the EU AI Act. The key requirements include:
Transparency and traceability: AI models must be explainable and comprehensible.
Regular validation: Models must be continuously reviewed and validated.
Governance: A governance framework is required that ensures the independent validation of the models.
Data quality: The data used must be of high quality.
Proportionality: The requirements for the models must be in proportion to their risk potential.
The authors highlight the specific challenges associated with the use of AI:
Complexity of models: AI models are often difficult to understand and require extensive resources for validation.
Expansion of the scope of application: AI is increasingly being used in areas such as customer advice and automated processes, which creates new risks.
Explainability and transparency: The high requirements of regulators for explainability are in conflict with the inherent complexity of many AI algorithms.
Continuous monitoring: Models must be continuously monitored and adapted to changing conditions.
Ethics: Banks must ensure that their models do not make discriminatory decisions.
Despite the challenges mentioned, the use of AI offers considerable opportunities. These include increased efficiency, improved decision-making, early risk detection and the potential for innovation, which should not be underestimated.
You can find the full position paper here.
Outlook for the year 2025
The Round Table plans to discuss the following questions in the coming year: How is GenAI being applied in the cyber domain and how does this change the potential threat? How can the trustworthiness of data for AI applications be increased? What criteria apply to a functional infrastructure (hardware (hosting, data centers) and software) in order to achieve greater efficiency in value creation? What requirements must be met with regard to policy, procedure and governance?
