When insurance companies calculate your premium, they're rarely using just the data you explicitly provide. A postcode might secretly stand in for ethnicity. An occupation might quietly proxy for sex. These hidden inferences—called proxy discrimination—can mean that customers are charged unfairly based on protected characteristics they never disclosed, buried beneath layers of algorithmic logic that even insurers themselves may not fully understand.
Now researchers at Bayes Business School, part of City University of London, have developed a framework to measure and reduce this invisible bias. The tool, published in the European Journal of Operational Research, offers insurance companies a concrete way to identify which variables in their pricing models contribute most to proxy discrimination—and, surprisingly, which ones can actually reduce it.
"The interplay between pricing factors and fairness is even more complex than previously recognized," says Professor Andreas Tsanakas, Professor of Risk Management at Bayes Business School and co-author of the research. The framework is applicable to most types of insurance cover and could be adopted across the sector, he notes. It could also extend beyond insurance into other financial services like credit scoring, where similar hidden biases can determine who gets approved and at what cost.
The real-world impact is concrete and urgent. When researchers applied their policyholder-specific measures to an actual motor insurance pricing model, they uncovered a troubling pattern: young drivers from one ethnic group were systematically quoted higher premiums. While the overall portfolio-level impact of proxy discrimination appeared minor, the harm concentrated on specific demographic groups was significant. That disparity was partially attributed to proxy effects—discrimination hiding in plain sight within the algorithm.
The challenge lies partly in incomplete information. Customers are often asked their sex, making that data available to insurers. Ethnicity, however, is rarely collected directly. Yet the framework developed by Tsanakas and his colleagues has begun addressing these gaps, finding ways to measure discrimination even when protected characteristics aren't explicitly recorded. The tool works by identifying which variables act as proxies and how strongly they contribute to unfair pricing outcomes.
What makes this framework particularly valuable is that it doesn't just diagnose the problem—it empowers action. Insurance companies can now use it to reduce indirect discrimination if they choose to. Auditors and regulators, too, can deploy it as a diagnostic tool. But as Tsanakas makes clear, the responsibility ultimately rests with policymakers. "It's up to regulators to set out clear principles and incentives for insurers to act on the issue," he says. Without clear regulatory pressure, insurers have little commercial motivation to unwrap their own algorithmic biases.
The research reveals something perhaps counterintuitive: some variables can actually reduce proxy discrimination rather than amplify it. That unexpected finding suggests that insurers have more levers to pull than previously thought—that fairness isn't simply about removing problematic variables, but about understanding the intricate relationships between all the factors that feed into a premium calculation.
For customers frustrated by opaque pricing and for regulators trying to protect vulnerable groups, this framework offers a path toward fairer insurance markets. The tool exists. The question now is whether the industry will use it.