Plaintiff win probability in US liability cases rose 20 to 30 percent between 2009 and 2024, and verdict awards more than doubled after 2020, in a May 2026 study by Tsz Chai Fung, Lie Ma, Liang Peng and Fang Yang (arXiv 2605.27265). Settlement amounts barely moved over the same period. A paid loss triangle contains both populations at once, which is why a single trend factor cannot see the shift that matters.

Key Takeaways

  • 20 to 30 percent higher plaintiff win probability from 2009 to 2024, measured by rolling-window logistic regression on case outcomes rather than inferred from development patterns.
  • Settlement probability fell more than 10 percent over the same period, so the mix of cases reaching verdict has itself moved adversely. The residual litigation pool is not a stable cross-section.
  • Settlement amounts showed limited and often statistically insignificant inflation while verdict awards more than doubled after 2020. Paid triangles mix the two in proportions that vary by accident year and lag.
  • 135 nuclear verdicts totalling $31.3 billion in 2024, up 52 percent year over year, with the median reaching $44 million against $21 million in 2020.
  • $15,000 in CAS prize money is pointed at exactly this problem, with the 2026 Reserves Call Paper Program naming LASSO, GAMs and Bayesian MCMC as methods for reflecting changes in loss cost and inflation trends.

What the Study Measured That a Development Factor Cannot

The headline is the 20 to 30 percent relative rise in plaintiff win probability. The usable part is the decomposition behind it: rolling-window logistic regression on case outcomes, quantile regression for severity, and random-weighted bootstrap for uncertainty.

That produces two findings a chain-ladder pattern cannot separate. Settlement probability declined more than 10 percent, meaning cases that stay in litigation skew toward the profiles that produce larger verdicts. And the inflation is concentrated: settlement amounts showed limited and often statistically insignificant movement across 2009 to 2024, while verdict awards more than doubled after 2020.

The effect is also heterogeneous. The study found social inflation more pronounced in corporate-defendant cases and in states without tort caps or third-party litigation funding regulation. A standard development factor estimated from the full portfolio aggregate applies one pattern to every accident-year tail with no attribution to the jurisdictional or defendant-profile composition of that year's open inventory.

Timing puts the finding next to a funded methodology search. The CAS 2026 Reserves Call Paper Program closed its submission window on June 15, 2026, with $15,000 in prizes for papers presented at the Casualty Loss Reserve Seminar in September, and its named scope covers using LASSO, Generalized Additive Models or Bayesian MCMC to reflect changes in loss cost and inflation trends.

One Factor Averaging Two Populations

The reserving consequence follows directly from where the inflation sits. A paid triangle blends settled cases, which carry limited social inflation exposure, with litigated-to-verdict cases, where the entire structural shift concentrates. Settlements are the numerically dominant resolution category, so an aggregate paid LDF understates social inflation's contribution to development velocity by construction, not by misapplication.

Splitting it takes three components rather than one trend selection:

  • Litigation frequency on an accident-year cohort: the share of claims entering suit rather than resolving without it.
  • Plaintiff win probability among litigated cases, the figure the study found has moved 20 to 30 percent.
  • Severity conditional on a plaintiff verdict, the channel through which nuclear verdict inflation compounds the first two.

The severity component is the one that has moved fastest. In 2024, 135 cases produced nuclear verdicts totalling $31.3 billion, a 52 percent increase over the prior year, with the median reaching $44 million against $21 million in 2020 (Marsh). An LDF captures what developed in prior accident years. It does not capture that the probability distribution of verdict outcomes shifted during the observation window it was fitted to.

Built on components, the IBNR margin becomes falsifiable. A load for litigation frequency trend, a separate load for plaintiff win probability calibrated to the book's jurisdictional footprint, and a severity tail load for the changed verdict distribution each release or hold on the same observable data that set them.

A single undifferentiated trend factor offers no such sensitivity, and the scale is not marginal. The joint CAS and Insurance Information Institute study put social inflation's addition to US commercial auto liability claim costs at more than $20 billion between 2010 and 2019, before the post-2020 verdict acceleration began.

The Inputs Are Outside the Carrier, and the Exposure Is Correlated

Every component above needs data the loss triangle does not contain, and mostly data the carrier does not own.

Lex Machina's docket analytics tracked 17,654 insurance cases filed in federal district courts in 2023, and reported claimants prevailing more than 55 percent of the time in business liability cases that reached trial. Counting all outcomes including dismissals, defendants prevail three times more often. Those two rates describe different things, and only the trial-specific one drives reserve exposure on cases that reach verdict.

The leading indicators sit further out. Third-party litigation funding is forecast to exceed $30 billion as a US industry, and in jurisdictions with disclosure requirements the presence of funding in a case predicts trial persistence and verdict magnitude. Attorney advertising spend by jurisdiction correlates with represented-claim rates and escalation to suit years before the resulting litigation reaches a triangle. All of it is third-party data on a purchased cadence, which is a different provenance and a different audit position than the carrier's own paid history.

Then the exposure travels across lines in a way single-line analysis does not capture. Umbrella accumulates through the attachment point: as verdicts push more cases above primary limits, umbrella attachment frequency climbs even when primary frequency is flat. A commercial auto primary limit that covered virtually all verdicts in 2018 is pierced routinely where median nuclear verdicts reach $44 million, pulling umbrella layers into cases that would once have closed inside the primary policy. E&O and D&O run the same pattern at higher thresholds.

A carrier writing commercial auto, GL, umbrella and E&O on the same corporate account concentrations holds one correlated exposure priced four separate times, and the jurisdictional concentration rather than the line mix is what decides whether the aggregate IBNR is adequate.

Further Reading

Sources

  1. Fung, Ma, Peng, and Yang: Quantifying Social Inflation in Liability Insurance with Advanced Statistical Methods (arxiv 2605.27265, May 2026)
  2. CAS: 2026 Reserves Call Paper Program on Improved Methodologies and Technologies for Reserving (Casualty Actuarial Society, 2026)
  3. CAS: Casualty Loss Reserve Seminar 2026 (September 14-16, 2026)
  4. CAS and Insurance Information Institute: Social Inflation and Loss Development (2022)
  5. Lex Machina: 2024 Insurance Litigation Report (LexisNexis, 2024)
  6. Marsh: Social Inflation and Nuclear Verdicts (median nuclear verdict data)
  7. NAIC: Insurance Topics, Social Inflation
  8. INSURICA: Social Inflation and Nuclear Verdicts, Definitions, Trends, Facts, and Figures (third-party litigation funding market size)