The Correlation Risk Hiding Inside Biometric Underwriting Data

The Assumption Biometric Underwriting Never Tested

Biometric data has made individual underwriting more precise than it has ever been. Steps per day, heart rate patterns, and activity levels now sit alongside traditional factors as genuine, independent predictors of mortality risk.

That precision is real, and it is why adoption has grown across life and health reinsurance portfolios. But precision at the individual level and safety at the portfolio level are two different claims, and only one of them has actually been tested.

What Is Biometric Risk Correlation After a Population Event?

It is the tendency of biometric mortality signals to move together across large parts of a portfolio at once, following a shared shock, instead of moving independently the way they did across isolated individual cases.

A pricing model built on biometric data typically treats each policyholder's signal as its own independent data point. That is a reasonable assumption in normal conditions, where one person's activity level has nothing to do with another's.

A population-level event changes that. A pandemic, an economic shock, or a public health crisis can shift biometric signals for large cohorts of people at the same time, in the same direction, which is exactly what "correlated" means in an actuarial sense.

Why Does This Problem Hide Behind Portfolio Growth?

Because growth is driven by individual-level precision, and individual-level precision does not test correlation risk at all.

Biometric-informed portfolios grow because the underlying data genuinely improves individual risk assessment. Research on wearable sensor data confirms this directly: steps per day "can effectively segment mortality risk even after controlling for age, gender, smoking status and various health indicators."

That is a strong, validated individual-level signal. Nothing about validating it at the individual level says anything about what happens when the same signal moves for thousands of policyholders simultaneously, which is a completely separate statistical question that adoption growth does not answer on its own.

Why Doesn't Individual Validation Answer the Portfolio Question?

Because individual validation measures prediction accuracy for one person at a time, holding everyone else constant.

A model can be excellent at ranking individual mortality risk and still be wrong about how those risks move together under stress. Those are different properties of a model, and testing one does not test the other.

This is the same blind spot that shows up in other areas of quantitative risk modeling, where a model validated on historical, independent-looking data quietly assumes that independence will keep holding in the future.

What Evidence Shows Population Events Create Correlated Risk?

Pandemic mortality research shows deviation tied to shared, population-level factors that move together rather than independently.

Research on managing pandemic risk after COVID-19 identified correlated mortality drivers across populations, including "prevalence of cardiovascular diseases, gross domestic product, Gini index, vaccine uptake of the first two shots, and proportion of population living in urban areas." None of these factors are individual-level, and none of them move independently across a population, they shift together, at the same time, for entire cohorts at once.

That is precisely the kind of correlated movement a biometric pricing model built on individual independence is not designed to anticipate. If activity levels, heart rate patterns, or other biometric signals shift alongside these same population-level drivers during a shock, the correlation shows up directly inside the biometric data itself.

Why Is Insured-Life Mortality Different From General Population Mortality Here?

Full medical underwriting removes many at-risk groups from an insured population, but the remaining group can carry other correlated exposures that a shock activates differently.

The same pandemic research notes that "mortality on fully underwritten products is significantly lower than that of the general population, because full medical underwriting usually removes at-risk groups, such as people with underlying chronic diseases." That sounds protective, and largely is, for the specific risks underwriting was designed to screen for.

But the research also flags a paradox worth sitting with: "this same demographic is more prone to travel internationally and reside in densely populated urban centers, factors which can introduce different risks." A population healthier by traditional underwriting standards is not automatically a population insulated from every kind of correlated shock, particularly ones tied to mobility and density rather than baseline health.

A joint research effort between RGA, the Society of Actuaries, and LIMRA found that "COVID-19 levied unequal effects on mortality rates in insured lives versus the general population, with significant differences by age." That finding alone confirms the core point, insured and general populations do not move together uniformly during a shock, which means a model calibrated on general population correlation assumptions will not automatically be right for an insured book either.

Is This Risk Limited to Wearable-Informed Pricing Programs?

No, any underwriting or pricing approach that treats individual biometric signals as independent risk factors carries this exposure.

Wearables are the most visible source of biometric data today, but the same correlation risk applies to wellness program screenings, periodic biometric health checks, and any other structured biometric input used to price or underwrite risk. The exposure is not about the data source, it is about the independence assumption baked into how that data gets used.

Reinsurers relying on tools like a Behavioral Biometrics Risk AI Agent as part of a broader risk assessment stack should confirm whether correlation across cohorts has been explicitly tested, not just assumed away by default.

How Is This Different From Traditional Catastrophe Mortality Risk?

Traditional catastrophe risk is modeled explicitly as a correlated event, while biometric-informed individual scoring is typically built and validated as if policyholders are independent.

Reinsurers already have decades of practice pricing correlated catastrophe risk, pandemics, natural disasters, mass-casualty events. That modeling tradition explicitly starts from the assumption that many lives can be affected together.

Biometric-informed underwriting grew up in a completely different tradition, one built around improving individual-level precision, where independence across policyholders was a reasonable working assumption most of the time. The gap this creates is structural, the discipline that knows how to model correlation and the discipline that built the biometric pricing tools have not, in most organizations, actually talked to each other about this specific risk.

What Is the First Step Toward Recognizing This Exposure?

Reviewing whether current biometric-informed pricing and underwriting models have ever been explicitly tested for correlated deterioration across cohorts following a shared population-level event.

For most reinsurers, the honest answer to that question right now is no, not because anyone made a deliberate decision to skip it, but because the question was never framed as something requiring a separate test. Individual-level validation was treated as sufficient, and portfolio-level correlation testing was never added as its own distinct step.

That gap is exactly what turns a well-validated individual pricing tool into an unpriced portfolio risk the moment a population event actually occurs. The margin and capital consequences of that gap are covered directly in the portfolio-profitability distortion created by biometric risk correlation after population events, and the underlying evidence-quality parallel is explored in why underwriting evidence ages too quickly for reinsurers to trust.

What Other Industries Have Faced This Same Correlation Blind Spot?

Credit risk modeling faced an almost identical blind spot before the 2008 financial crisis, when individually well-validated default models failed to account for correlated default across the portfolio.

Individual mortgage default models at the time were genuinely good at predicting whether a specific borrower would default, based on that borrower's credit history, income, and loan characteristics. What those models had never been meaningfully tested against was what happens when a shared macroeconomic shock causes defaults to move together across large numbers of borrowers simultaneously.

That is structurally the same gap described in this post, individual-level accuracy mistaken for portfolio-level safety. Reinsurers do not need to repeat that lesson from scratch, the parallel is close enough that the correlation-testing discipline built in credit risk modeling after 2008 is directly relevant to how biometric mortality correlation should be approached now, before rather than after a comparable event forces the question.

How Fast Can Correlation Actually Develop Once a Population Event Starts?

Correlation in biometric signals can develop within weeks of a population-level event, well before it would be visible in traditional lagging mortality data.

Biometric signals like activity levels and heart rate patterns are collected continuously, which means they can reflect a population-wide behavioral or physiological shift far faster than claims data or mortality statistics, which typically lag the underlying event by months or years. That speed is actually an opportunity as much as a risk, a portfolio with real-time biometric correlation monitoring could, in principle, detect a developing correlated shock earlier than any traditional mortality-tracking metric would.

Realizing that opportunity requires the monitoring capability to already be in place before the event happens. A reinsurer building this monitoring only after a shock has started is working from the same lagging position traditional mortality tracking already has, losing exactly the early-warning advantage biometric data is uniquely positioned to provide.

Biometric data earned its place in modern underwriting because it works, individually, and that is not in question. What remains untested, for most portfolios, is what happens when the same signal that made one policyholder's pricing more accurate starts moving in the same direction for thousands of policyholders at once.

Sources

• RGA: Managing Pandemic Risk After COVID-19: Lessons from the past, preparing for the future
• Munich Re: Physical activity measured by wearable sensors can effectively assess mortality risk