The most recent practice resource document published for consultation by the Canadian Institute of Actuaries is focused on bias and fairness in actuarial work. The work is challenging, they point out, as even artificial intelligence ethics scholars say that “fairness is dynamic and social and not a statistical issue.”
That said, they add that this does not mean that it cannot be a dimension of actuarial model evaluation.
“This paper is not intended to serve as a step-by-step guide outlining exactly how to address every possible fairness issue that may arise in a rating algorithm. The context of an individual situation is important to consider in practice,” they write. “This paper serves as a starting point for the practitioner to better understand concepts of bias and fairness without attempting to be exhaustive on these topics.” (The paper also includes an extensive list of resources.)
They point out that multiple definitions of bias are available, and no single definition of fairness exists. “Like other concepts, it is constantly being debated and adapted in a democracy.”
Those measuring bias in the property and casualty (P&C) underwriting lifecycle can look at multiple sources, including improper data collection, models and assumptions. “All of these sources should be considered. Many measurement approaches are possible and these can then be used to test for the presence of bias in outcomes or results,” they write. “Bias assessment could also form part of a broader internal model governance plan.”
The paper’s audience include P&C pricing actuaries and those working in other areas of actuarial practice.