Actuarial Notes

Bias and Fairness in Pricing and Underwriting of Property and Casualty (P&C) Risks

By Canadian Institute of Actuaries · 2023 · 223056

Study guide: Exam 6C

A CIA practice resource document giving P&C pricing actuaries tools for detecting, evaluating and mitigating potential bias in actuarial risk assessment models. It was published in April 2023 (document 223056) by the Bias and Ethics Subcommittee of the Committee on Property and Casualty Insurance Pricing, and its recommendations are, in its own words, "not prescriptive or binding".

On the syllabus

  • Exam 6C — objective A1; Sections 1, 2 and 3.

Executive summary

  • Bias is defined as a measurable property of predictive models; fairness concerns the use of those models to make decisions with consequences for people.
  • The concepts are closely linked with Section 1400 of the Standards of Practice.

1 Intent, scope and cross-references

  • 1.1 Intent
    • Growing reliance on automated processes, models and machine learning to set premiums can let unwanted bias into the practitioner's work product; the paper focuses more on bias than on fairness.
  • 1.2 Scope
    • Tools and considerations for developing risk segmentation or tiers, measuring price differentials, discounts and surcharges, and other predictive analytics; it does not apply to the societal determination of notions of fairness.
  • 1.3 Cross-references

2 Historical issues and current evolution

  • Fairness is a social construct that evolves over time and across societal contexts, and can be assessed through quantitative measurements.
  • Illustrations: a correctional risk-assessment score found to be biased against Indigenous and Black inmates; redlining; and the use of gender in auto rating, which the Supreme Court of Canada allowed in Zurich Insurance Co. v. Ontario (Human Rights Commission), with two justices dissenting.

3 Definitions

  • 3.1 Bias
    • Bias in P&C pricing is any situation in which the outcomes of ratemaking models are systematically less favourable to individuals within a particular group and no relevant difference between groups justifies the difference in premiums or rates.
    • This is not the statistical definition (an estimator whose expected value differs from the true value), which the paper calls an unrelated concept.
  • 3.2 Direct and indirect discrimination
    • A pricing model avoids direct discrimination if no characteristic protected by human rights legislation is used as a rating factor; it avoids indirect discrimination if, in addition, the other features do not allow the protected ones to be inferred.
  • 3.3 Fairness
    • No single definition or measure of fairness exists; the practitioner weighs who is harmed, how much, how many, whether the product is essential, and whether society views the price discrimination as unfair.
    • Fairness splits into procedural (how insureds are treated through the pricing process) and distributive (how pricing outcomes are distributed across insureds).
  • 3.4 Contrasting bias and fairness
    • Bias does not necessarily imply unfairness, and a lack of fairness does not necessarily imply bias; bias is static, while an outcome that is fair today may not be seen as fair in the future.
  • 3.5 Ethics
    • Practitioners have an unequivocal obligation to uphold the law, including protected-class legislation that varies by province (for example, auto insurers may not use credit information in Ontario).

4 Analysis and measurement of issues

  • 4.1 Overview
  • 4.2 Measurements
    • 4.2.1 Measurement of bias
    • 4.2.2 Purpose of measurement
    • 4.2.3 Other relevant considerations
  • 4.3 Ethical framework

5 Next steps for the practitioner

  • 5.1 Bias and Fairness Assessment Documentation
    • 5.1.1 Framework
    • 5.1.2 Documentation considerations
  • 5.2 Reporting within model governance

6 Further research

7 Illustrative example

Appendix 1 Various definitions we rely on

Appendix 2 Examples of ethical frameworks

Appendix 3 Generic definitions of bias

Cited References

Related readings

  • CIA CSOP — the Standards of Practice, whose Section 1400 the paper says its concepts are closely linked with
  • Basic Ratemaking (Werner - 2016) — cited in §1.1 for a premium as a rate per unit of exposure

Sources