Verdicts Are Not Payouts

TL;DR — Raw verdict data is heavily skewed by selection bias and rarely reflects final payouts due to appeals and policy limits. Accurate forecasting requires isolating verdict data from settlement data to generate calibrated, honest settlement ranges rather than inflated point estimates.

A jury reads a nine-figure verdict. The plaintiff attorney issues a press release. The industry panics over social inflation. But the defense does not cut a check for that amount. The final payout is negotiated months later. It is constrained by policy limits, appellate risk, and the simple reality of collectability. A verdict is a headline. A payout is a transaction. Yet when legal data vendors build predictive analytics, they rely heavily on verdict data. Verdicts are public and easy to scrape. Settlements are buried in private corporate databases. This data availability problem creates a severe structural flaw in insurance forecasting. If you train a machine-learning model to predict case value using raw verdict data, you are training it on a fiction. You teach the model that outlier jury awards are the expected outcome. You guarantee a massive upward bias in your predictions, ensuring your early reserves will be wildly inaccurate.

This bias is a well-documented phenomenon in legal economics. The Priest-Klein hypothesis demonstrates that cases reaching a verdict are never representative of the underlying claims. They are statistical anomalies. They represent a failure of the settlement market. Cases go to trial when the plaintiff and the defense have mutually exclusive, highly confident estimates of the outcome, often driven by asymmetric information. The vast majority of cases that settle before trial belong to an entirely different distribution. Even when a massive verdict is rendered, research by legal scholars shows a vast disconnect between the jury award and the final payout. A fifty-million-dollar award against a defendant with a five-million-dollar policy limit is a five-million-dollar settlement waiting to happen. The distribution of actual payouts is truncated. Building a reserving strategy on raw, untruncated verdict data is mathematically illiterate.

Neural-symbolic boundaries

This structural reality explains why throwing a large language model at a pile of legal dockets is a mistake. Generative AI is an engine for text prediction, not mathematical forecasting. It is built to generate the next plausible word based on its training data. It collapses distinct probability distributions into a single, blurry output. If its training data is saturated with high-profile verdict press releases, the model will generate inflated settlement estimates. It lacks the mathematical architecture to distinguish between a public headline and a confidential payout. We enforce a strict boundary between reading and forecasting. At Canotera, generative AI reads the case file. It ingests thousands of pages of pleadings, medicals, and correspondence. It maps the unstructured text into a structured, symbolic representation of the facts. It does the reading and the structuring. It does not do the prediction.

The actual prediction happens in a separate mathematical system. We use geometric machine-learning models trained strictly on large numbers of resolved cases with known outcomes. To prevent upward bias, we isolate the data. Verdict-heavy comparables are kept in a separate module from settlement-anchored data. Mixing them poisons the well. You cannot triangulate a realistic settlement by averaging it with a runaway jury award. By keeping these distributions distinct, the models learn the actual mechanics of resolution. They learn how cases settle when liability is uncertain, when third-party litigation funding is involved, or when venue-specific priors shift the baseline.

The math of the settlement haircut

Many claims organizations attempt to solve the verdict bias manually. They apply a heuristic haircut to jury awards. They look at a nuclear verdict and arbitrarily slash it by a fixed percentage to estimate a settlement value. This approach assumes a linear relationship between a jury award and a negotiated settlement. That relationship does not exist. The discount applied to a verdict depends heavily on appellate asymmetry, liability probability, and simple collectability. A verdict against an insolvent defendant has a cash value of zero, regardless of what the jury foreperson read into the record. Beyond collectability, the defense has a strong incentive to appeal a nuclear verdict. A plaintiff who wins a massive verdict faces years of appellate litigation, during which they collect nothing. The time value of money and the risk of a total reversal incentivize them to accept a fraction of the award immediately. That appellate risk forces a settlement bounded by the probability of reversal, not by the raw jury number. A fixed haircut ignores this entirely.

Forecasting litigated claims requires honest error reporting. A single point estimate is useless because it masks the underlying variance. Our models output a calibrated settlement range. Calibration means that our stated confidence matches reality. If the model predicts a settlement range with 80 percent confidence, the actual payout falls within that range exactly 80 percent of the time. We use conformal ranges rather than traditional confidence intervals because conformal prediction does not force assumptions about the shape of the data distribution. It guarantees mathematical coverage even when the claims data is noisy and non-parametric. It tells the claims executive exactly how wide the uncertainty is on day one. It provides a realistic bound for setting reserves and allocating defense spend.

Every output is traceable. When the platform produces a reserve delta versus the current reserve, or flags an escalation probability, it provides the specific drivers behind those numbers. The mathematics link directly back to the source documents extracted by the generative layer. A claims professional can see exactly which medical records or jurisdiction histories are driving the range. The point is to negotiate from data. When a plaintiff attorney anchors a demand to an irrelevant nuclear verdict from a neighboring county, the defense needs a mathematical counterweight. You cannot manage reserve volatility with gut instincts or models trained on public relations. If your baseline is built on jury awards, you are reserving for the headline. We reserve for the check.