# Mitigating Discrimination in Insurance with Wasserstein Barycenters

**Authors:** Arthur Charpentier, François Hu, Philipp Ratz  
**arXiv:** [2306.12912](https://arxiv.org/abs/2306.12912)  
**Date:** 2023-06-22  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2306.12912v1/

## Key points

- The insurance industry heavily relies on predictive models to assess risks based on customer characteristics
- Predictive models have faced criticism for perpetuating discrimination based on sensitive features like gender or race
- Discrimination is often a result of historical data biases
- Simply excluding sensitive variables in pricing process is ineffective in addressing discrimination
- Authors propose using Wasserstein barycenters instead of simple scaling to mitigate biases
- Wasserstein barycenters consider distributional differences between groups, offering a more effective approach to mitigating discrimination
- Authors demonstrate the effects and effectiveness of this approach through real data analysis and discussion of implications
- Lars Powell's work highlights importance of accurately assessing risks in insurance pricing
- Rebecca Rhynhart's study shows how historical biases can affect discriminatory practices
- Jonathan J Rolison et al.'s research emphasizes need for comprehensive analysis when addressing complex issues like discrimination in insurance
- Walter Rudin's book provides foundational knowledge for statistical modeling techniques used in this study
- Daniel Sabbagh's work offers insights into affirmative action strategies relevant to addressing discrimination in insurance practices
- Filippo Santambrogio's book provides mathematical concepts related to transportation problems applicable to understanding bias mitigation techniques
- The Zebra's report offers practical insights into how different variables are currently considered in insurance pricing processes by state
- Yves Thiery and Caroline Van Schoubroeck's research provides broader perspective on challenges of achieving fairness in insurance practices
- Ronald Turner's article discusses ethical considerations surrounding discriminatory practices

## Summary

The insurance industry heavily relies on predictive models to assess risks based on the characteristics of potential customers. However, these models have faced criticism for perpetuating discrimination based on sensitive features such as gender or race. This discrimination is often a result of historical data biases. As the industry shifts towards machine learning-based predictions, there is a growing call for greater mitigation of biases. Simply excluding sensitive variables in the pricing process has been shown to be ineffective in addressing discrimination. In this article, the authors investigate why predictions are necessary in the insurance industry and why correcting biases is not as straightforward as identifying a sensitive variable. They propose using Wasserstein barycenters instead of simple scaling to ease biases. Wasserstein barycenters offer a more effective approach to mitigating discrimination by considering the distributional differences between groups rather than just scaling variables. The authors demonstrate the effects and effectiveness of this approach by applying it to real data and discussing its implications. To provide further context, the authors reference several relevant sources. Lars Powell's work on risk-based pricing of property and liability insurance highlights the importance of accurately assessing risks in insurance pricing. Rebecca Rhynhart's study on mapping the legacy of structural racism in Philadelphia sheds light on how historical biases can affect discriminatory practices. Jonathan J Rolison et al. 's research on factors contributing to road accidents emphasizes the need for comprehensive analysis when addressing complex issues like discrimination in insurance. Walter Rudin's book "Real and Complex Analysis" provides foundational knowledge that may inform statistical modeling techniques used in this study. Daniel Sabbagh's work on equality and transparency offers insights into affirmative action strategies that could be relevant to addressing discrimination in insurance practices. Filippo Santambrogio's book "Optimal Transport" provides mathematical concepts related to transportation problems that may be applicable to understanding bias mitigation techniques. The Zebra's report on car insurance rating factors by state offers practical insights into how different variables are currently considered in insurance pricing processes. Yves Thiery and Caroline Van Schoubroeck's research on fairness and equality in insurance classification provides a broader perspective on the challenges of achieving fairness in insurance practices. Finally, Ronald Turner's article on stopping discrimination based on race offers additional insights into the ethical considerations surrounding discriminatory practices. Overall, this article aims to contribute to the ongoing discussion on mitigating discrimination in the insurance industry by proposing a novel approach using Wasserstein barycenters and providing empirical evidence of its effectiveness.