Introduction

In the dynamic landscape of financial regulation, the accuracy and reliability of internal rules-based (IRB) models for regulatory capital calculations have become increasingly critical.  The Treasury’s Basel IV endgame reproposal presents a valuable opportunity to assess the application of artificial intelligence (AI) to these IRB models, enhancing supervisory oversight and ensuring system integrity.  A September 2024 article written by two Duke University professors (the “Duke article”)^[1] proposed a four-part test for evaluating artificial intelligence (AI) models. The Center for the Study of Financial Market Evolution (CSFME) supports incorporating this four-part test, adapted for the context of IRB models, into the Federal Reserve Board’s (FRB) supervisory practices. This rigorous methodology, designed to assess the robustness and interpretability of AI models, can provide a valuable tool for FRB examiners reviewing internal models. By applying this framework, the FRB can promote greater transparency, enhance risk management capabilities, and foster trust in using AI within the financial services sector.

The Four-Part Test

The authors of the Duke article argue that the use of AI by governments in regulation and law enforcement should undergo independent assessments. These evaluations should focus on the quality of the training data, the testing of that data, the predictive model itself, and its application in specific cases. They also emphasize the need for interpretable AI, allowing people to understand how it reaches results in a particular case. The article highlights that AI systems must be tested to ensure they perform as promised in high-impact settings and the authors propose a four-part framework for assessing the reliability, interpretability, and transparency of AI models.

1. Independent Performance Testing. Outsiders should independently assess AI models without a financial motive. This assessment should evaluate the quality of training data, test data, the predictive model, and its application to specific cases.
2. Availability of Code and Data for Testing. The code and data used to develop the AI model should be shared for outside testing. For sensitive data, sharing should be subject to restrictions or court orders.
3. Disclosure of Test Results. The results of AI model tests should be fully disclosed. This disclosure will allow for transparency and understanding of the model’s strengths and limitations.
4. Interpretability of AI. AI models should be interpretable, enabling users to understand how the model reaches results in individual cases.

Application of the Four-Part Test by FRB Supervisors

The Federal Reserve Board (FRB) should incorporate this four-part test into its supervisory framework for reviewing IRB models used for regulatory capital calculations.

Currently, banks can utilize internal models but must adhere to the standardized approach if it results in higher capital requirements. The proposed four-part test would add another layer of scrutiny, focused on the reliability and interpretability of AI-driven IRB models.

Benefits of Applying the Four-Part Test

• Enhanced Risk Management: A rigorous testing regime can identify potential weaknesses and biases in AI models, leading to more robust risk management practices.
• Improved Transparency: Sharing code, data, and test results promotes transparency and allows for independent verification of model accuracy.
• Greater Trust and Confidence: A clear understanding of how AI models function fosters trust among regulators, financial institutions, and the public.
• Encouragement of Innovation: A well-defined testing framework can encourage innovation by providing clear guidelines for developing and deploying AI models.

Addressing Potential Concerns

• Industry Pushback: Some financial institutions may resist the proposed testing requirements, citing concerns about proprietary information or compliance costs. However, the potential benefits of enhanced risk management and transparency outweigh these concerns.
• Regulatory Overreach: It is crucial to find a balance between ensuring responsible AI use and avoiding overly burdensome regulations that could hinder innovation. The four-part test presents a comprehensive yet adaptable framework for assessing different AI models and applications.

Implementation Considerations

• Phased Implementation: A phased implementation of the testing requirements could allow financial institutions time to adapt and ensure compliance.
• Collaboration and Guidance: The FRB should collaborate with industry stakeholders and provide clear guidance on the testing process and expectations.
• Ongoing Monitoring and Evaluation: The FRB should continuously monitor the effectiveness of the testing regime and make adjustments as needed to ensure it remains relevant and practical.

Conclusion

Adopting AI in financial services presents a unique opportunity to enhance efficiency and manage risk. However, ensuring the responsible and transparent use of AI models is crucial. The proposed four-part test, adapted from the Duke article, provides a robust framework for evaluating deep learning models. By incorporating this test into their supervisory practices, FRB examiners can promote the responsible adoption of AI in financial services while ensuring the stability and integrity of the financial system.

In response to the Treasury’s RFI on AI in financial services, CSFME proposed a research project to evaluate deep learning models for risk management and forecasting within the context of securities finance. The proposed research emphasizes the importance of transparency, ethical AI use, and developing supervisory frameworks for responsible AI adoption. CSFME is committed to enhancing financial stability and regulatory compliance through advanced AI methodologies, data-driven insights, and collaboration with the Treasury.

[1] Garrett, Brandon L. and Rudin, Cynthia, Testing AI (September 06, 2024). Duke Law School Public Law & Legal Theory Series No. 2024-62, Available at SSRN: https://ssrn.com/abstract=4948789 or http://dx.doi.org/10.2139/ssrn.4948789