The CAS Statement of Principles Regarding Property and Casualty Insurance Ratemaking is a foundational document for any pricing actuary. Its introduction emphasizes the critical role of employing proper actuarial procedures to derive rates that “protect the insurance system’s financial soundness and promote equity and availability for insurance customers.”
This involves compliance with four criteria for actuarially sound rates:
•Reasonable
•Not excessive
•Not inadequate
•Not unfairly discriminatory
The CAS has committed to providing industry-leading research to balance these principles within insurance pricing applications. This commitment led to the Series on Race and Insurance Pricing, a group of four CAS Research Papers released in 2022. Now, the CAS and its many volunteer members are preparing to release a Phase II of the series in 2024.
The recent CAS Spring Meeting included a concurrent session providing an overview of the latest race and insurance pricing series before going deeper into two of the papers. CAS Fellow Mallika Bender, diversity, equity & inclusion staff actuary at the CAS, moderated the session and provided an overview of the series. She noted that these papers reflect the latest understanding of race and algorithmic bias, but that regulator and consumer perspectives on these topics are continually evolving.
Scott Merkord, consulting actuary at Risk & Regulatory Consulting LLC (RRC), and Rich Moncher, senior consultant at Octagram Analytics, provided an overview of the research into regulatory perspectives on algorithmic bias and how actuaries can prepare. Merkord and his RRC colleagues conducted the survey of regulators and will be authoring the first paper. Moncher and his Octagram colleagues will be writing the second paper.
Recent regulatory and legislative actions in the United States are emanating from the NAIC and individual states, as well as the federal government. In 2020, the NAIC formed its Special Committee on Race and Insurance, culminating in the issuance of its “Model Bulletin regarding the Use of AI Systems by Insurers” in 2023. They have also issued model review questionnaires and have the Cybersecurity and Technology (H) Committee continuing to focus on all aspects of bias and insurance.
At the time of their presentation, six states (Connecticut, Hawaii, Illinois, New Hampshire, Rhode Island and Vermont) have adopted the NAIC’s Model Bulletin. Many other states are actively considering adoption, so this list is expected to grow throughout 2024. Three other states (California, Colorado and New York) have introduced their own regulations.
State regulators and/or legislatures are also taking action. Colorado passed its Senate Bill 21-169 in 2021 addressing unfair discrimination of various protected classes. Their insurance department has since issued guidance for the life insurance industry and is now working on comparable guidance for P&C products. So far, no other state has adopted a “copycat” of Colorado’s legislation.
California’s insurance commissioner issued Bulletin 202205 in 2022, setting fairness principles to avoid bias in marketing, policy issuance, pricing, fraud investigation and claims handling. The Connecticut Department issued “Big Data and Avoidance of Discriminatory Practices” in 2022. The District of Columbia Department formed its Diversity and Equity Inclusion Committee to explore how to regulate issues of bias in insurance. The New York Department released “Use of AI Systems and ECDIS [External Consumer Data and Information Sources] in Insurance Underwriting and Pricing” in January 2024, while Illinois legislators have introduced bills to dramatically restrict the use of many variables for personal auto on the grounds of unfair discrimination.
U.S. insurers also need to consider the maze of federal civil rights acts. Then they’ll need to layer on relevant model acts and laws, such as Unfair Trade Practices Model Act, Unfair Claims Settlement Practices Model Act and Property & Casualty Model Rating Law.
With the table now set, Merkord provided an overview of the underlying research conducted for the “Survey of Regulators on Algorithmic Bias” paper. Risk & Regulatory Consulting sent their survey to all state insurance departments and received responses from ten states. The survey questions focused on three areas: responsibility of insurers with regard to algorithmic bias in their models, concern regarding rating elements utilized in private passenger auto (PPA) and regulatory perspectives on algorithmic bias.
The paper will go far deeper on all their conclusions from the regulator survey, but I found four takeaways to be most interesting.
• Most respondents agree that multiple testing methodologies should be used to identify algorithmic bias
• Respondents are mixed on whether race should be used to test for algorithmic bias, and many disagree with the use of Bayesian Improved First Name Surname and Geocoding (BIFSG) as a technique to infer race.
• Regarding PPA rating elements, most respondents are concerned about the use of homeownership, occupation, credit-based insurance scores and criminal history.
• Most respondents do not believe evaluating rates for actuarial soundness alone satisfies their concerns surrounding unfair discrimination.
Moncher provided an overview of Octagram’s paper entitled “Approaches to Respond to Bias Regulation.” An opening key question is: “During which stage(s) of the modeling process should bias and fairness be considered?” Options include project planning, data preparation and exploration, model training, model evaluation and selection, and model implementation.
I will jump to the punchline by sharing their recommendation to consider bias and fairness in every stage of the modeling process. Their paper will, and Moncher’s presentation did, walk through many ways in which bias should be considered at every stage, regardless of the model application. More poignantly, if it’s not considered during project planning, then the organization risks inefficiencies in time and effort by having to loop back to earlier stages of the project when bias is eventually considered.
In closing, Bender encouraged everyone to watch CAS communications for the release of all five research papers. The fifth piece is intended to be a handbook to help practicing actuaries apply new bias measurement and mitigation techniques into their actuarial work. This will help our profession deliver the highest standard of care to the policyholders whom the insurance industry serves every day, consistent with our CAS Statement of Principles.
Editor’s note: The sixth and final Research Paper in the series will compare international regulations on bias in AI. ●
Dale Porfilio, FCAS, MAAA, is the chief insurance officer at the Insurance Information Institute and president of the Insurance Research Council.
