Beyond Expected Cost: Considering Volatility Metrics in Stop-Loss Underwriting

Consider two employer groups with nearly identical demographic and geographic mixes, network discounts and stable large claimant history. Both have similar projected per member per month claims costs and pass underwriting review. Both are bound as profitable risks. A year later, one has performed within expected range; the other has generated a loss ratio of 172%. Same mean but profoundly different outcomes.

This is not an unusual scenario in stop-loss underwriting—it is a structural feature of the product. And it highlights a potential limitation of underwriting approaches that focus on expected cost.

Stop-Loss Performance Reflects Both Mean Cost and Volatility

Stop-loss underwriting has long centered on expected cost estimation: demographics, experience rating, credibility weighting and trend adjustment. These tools are well developed and necessary. But they are designed primarily to estimate central tendency—the mean of a cost distribution. Stop-loss indemnification, by contrast, is triggered by deviations from that mean. The product’s financial performance can be materially affected by variance and tail risk, in addition to expected cost.

Two groups with similar projected costs may exhibit materially different distributional characteristics: different tail thickness, different catastrophic claim clustering, different concentrations of specialty pharmaceutical exposure, different probability mass[1] in high-cost percentiles. Traditional underwriting inputs may capture the mean with reasonable accuracy while leaving dispersion largely unmeasured. Joshua W. Axene, FSA, FCA, MAAA, a partner and consulting actuary at Axene Health Partners, notes, “For stop-loss portfolios, variance relative to premium can be more predictive of financial performance than the expected claim cost. Elevated dispersion increases the risk of adverse deviation, even when the portfolio mean is actuarially sound.”

Retrospective Evidence: Variance Concentration in Practice

A retrospective actuarial analysis of a leading US stop-loss portfolio comprising 19 employer groups and 16,823 covered members examined whether pre-bind dispersion signals predicted ex post performance.[2] All groups had been quoted, bound and assumed profitable at the time of underwriting. Risk metrics were generated by Merit Medicine using its Merit Predict platform—an AI-led predictive analytics framework that produces member-level medical and pharmacy spend projections, a Stop-Loss Risk Score (conceptually a coefficient of variation) and an Aggregate Risk Score (expected first-dollar spend relative to a national benchmark)—based exclusively on information available at or before the underwriting decision, with no post-bind claims experience used as input. The underlying analysis was independently validated by Joshua Axene.

As with any retrospective analysis of this size, results are directionally indicative rather than statistically conclusive. The 19-group sample is sufficient to illustrate variance concentration as an actuarial phenomenon but is not a credibility-weighted basis for parameterizing pricing models. Broader validation across larger and more heterogeneous portfolios would be required before treating dispersion stratification as a primary pricing input.

The analysis produced two group-level measures. The first, the Aggregate Risk Metric, expresses a group's expected first-dollar healthcare spend relative to a national benchmark, where a score of 1.0 represents the benchmark mean, scores above 1.0 indicate higher-than-average expected spend and scores below 1.0 indicate lower-than-average expected spend. It is a mean-focused measure consistent with traditional underwriting. The second, the Stop-Loss Risk Metric, captures volatility of predicted member-level cost outcomes within the group. It is conceptually aligned with a coefficient-of-variation framework: lower values indicate stable, predictable claim patterns, while higher values indicate greater volatility and elevated stop-loss exposure.

Groups were then stratified into five tiers, from Tier 1 (lowest dispersion, or stable risk) to Tier 5 (highest dispersion, or extreme tail risk). The portfolio results were striking in their concentration. Six of 19 groups generated 67.4% of total reimbursements. Four groups classified in the highest-dispersion tier (Tier 5) accounted for 52% of underwriting losses before expenses. Tier 5 groups produced a combined loss ratio of 171.9%, compared to 55% for the remaining groups in Tiers 1 through 4. Two individual Tier 5 groups experienced reimbursements 172% greater than premium collected.[3]

From an actuarial standpoint, the results are consistent with the possibility that distributional risk was not fully captured in underwriting rather than any mean costs.

Modeling the Counterfactual

To assess the practical significance of dispersion stratification, a counterfactual underwriting scenario was modeled under the assumption that highest-dispersion groups would have been either declined or repriced to reflect their volatility exposure. The modeled impact was substantial. The scenario assumed two underwriting actions for highest-dispersion (Tier 5) groups: either declining to quote or applying a premium load sufficient to reflect their volatility exposure (which, in a competitive bid environment, would typically result in the bid being lost). All other groups were held at their actual bound terms. Under those assumptions, the four Tier 5 groups—which together had generated approximately $3.5 million in pre-expense underwriting losses—were removed from the realized portfolio results.

Portfolio loss ratio declined from 83.6% to approximately 55%. Underwriting margin improved by 107%. Reimbursements declined by approximately 50%.

Importantly, these improvements do not arise from eliminating average risk from the portfolio. They arise from isolating extreme dispersion. The modeled scenario suggests that better identification of high-variance risks could be associated with improved portfolio stability and more accurate stop-loss pricing.

Capital Implications of Unidentified Dispersion

Stop-loss underwriting volatility has downstream implications beyond the line of business itself. High-dispersion groups increase the probability of capital-consuming outcomes. Where high-dispersion groups are not identified, insurers may respond with broader conservatism. Alternatively, more granular risk stratification may improve capital efficiency in some settings. As Joseph Axene explains, “Concentrated variance carries meaningful capital consequences. If dispersion is not explicitly modeled within pricing and risk frameworks, required capital effectively backstops unmeasured volatility. Improved risk stratification enhances capital efficiency and reduces unintended capital drag[4].”

More precise dispersion measurement supports improved stress testing, more accurate tail quantile estimation, refined reinsurance strategy evaluation and stronger treaty negotiation positioning—all areas where granular risk characterization translates directly into economic value.

Operationalizing Volatility Metrics

Introducing dispersion metrics into actuarial and underwriting workflows is not purely a technical exercise. Considerations for incorporating this new technique often include explainability, auditability, governance and back-testing.

Perhaps most importantly, dispersion metrics must be understood as probability instruments, not deterministic predictors. They estimate the likelihood of adverse deviation, not its certainty. Treated as decision-support tools, they improve visibility into risk structure while preserving the actuarial judgment that remains essential to sound underwriting.

A Profession-Level Consideration

The actuarial conditions that make dispersion measurement valuable are not static, rather they are intensifying. The increasing prevalence of high-cost specialty pharmaceuticals, gene and cell therapies and complex oncology protocols is expanding the range of outcomes possible within small and midsize employer groups. As the right tail of the cost distribution grows heavier and less predictable, mean-only underwriting becomes progressively less sufficient as a risk characterization tool.

This is not a novel concept in insurance actuarial practice. Variance modeling has long been embedded in property-casualty pricing frameworks, where tail risk and catastrophic loss potential are explicitly parameterized. Stop-loss, which shares many of these structural characteristics, may be approaching a similar evolution.

For actuaries working in stop-loss pricing, managing general underwriter (MGU) oversight, captive funding strategy and reinsurance structuring, explicit incorporation of volatility metrics may represent an emerging area of practice development and, in time, an expected component of disciplined underwriting.

Conclusion

The retrospective portfolio examined here demonstrated what experienced stop-loss actuaries have long understood intuitively: adverse outcomes concentrate among a small subset of high-variance groups in ways that expected cost estimation alone does not predict. Four of 19 groups drove 52% of underwriting losses. A modeled counterfactual suggests that identifying those groups ex ante could have reduced the portfolio loss ratio by roughly 30% and improved margin by 107%.

The takeaway is structural, not technological. Expected cost estimation remains foundational. But it does not fully characterize stop-loss risk. Explicit measurement of dispersion—conceptually aligned with coefficient of variation and generated from pre-bind data—provides earlier visibility into tail exposure, supports more disciplined capital allocation and enables underwriting decisions that better reflect the product’s actual risk architecture.

As volatility within medical risk pools continues to grow, the actuarial toolkit for managing it may need to expand accordingly.

This article is provided for informational and educational purposes only. Neither the Society of Actuaries nor the respective authors’ employers make any endorsement, representation or guarantee with regard to any content, and disclaim any liability in connection with the use or misuse of any information provided herein. This article should not be construed as professional or financial advice. Statements of fact and opinions expressed herein are those of the individual authors and are not necessarily those of the Society of Actuaries or the respective authors’ employers.

Ali Panjwani is founder and CEO of Merit Medicine. Ali can be reached at ali@meritmedicine.com.

Endnotes

[1] Tail thickness is a measure of how much probability is concentrated in the extreme (low-frequency, high-severity) portion of a cost distribution. A “thicker” tail means catastrophic claims are more likely than a normal distribution would predict—a critical feature in stop-loss, where indemnification is triggered by exactly those extreme outcomes. Probability mass is the share of total probability assigned to a defined range of outcomes. In this context, “probability mass in high-cost percentiles” refers to the proportion of expected claims experience that sits in the upper percentiles of the cost distribution (e.g., the 95th percentile and above).

[2] Merit Medicine, “Using AI-Led Predictive Analytics to Improve Underwriting Outcomes,” White Paper, February 2026, https://www.meritmedicine.com/_files/ugd/1d68ef_a1b5d395a19145ab8ca0d7678f468a9e.pdf.

[3] Merit Medicine, “Using AI-Led Predictive Analytics to Improve Underwriting Outcomes.”

[4] Capital drag is the opportunity cost incurred when an insurer must hold additional risk-based capital to backstop unmeasured or imprecisely characterized volatility. Capital held against unidentified tail risk is unavailable for new business, investment or shareholder returns, reducing return on capital even when no adverse outcome materializes.