When Settlement Data Is Sparse, Triangulation Beats Imputation

TL;DR — Generative AI should extract case facts, but separate geometric models must calculate the settlement range using liability probabilities and policy limits. Mixing public verdicts with private settlements destroys model calibration.

If you build a predictive model for insurance litigation by training it solely on public docket data, you are training it on a statistically mutilated sample. The Priest-Klein hypothesis dictates that cases reaching a verdict are the anomalies. They are the boundary disputes where the plaintiff and the defense fundamentally miscalculate their respective odds of success due to asymmetric information or entrenched cognitive bias. The vast majority of claims settle privately. The data you actually need to set accurate reserves is hidden in closed files and confidentiality agreements, leaving a highly sparse matrix of known settlement values.

The standard data science response to a sparse matrix is imputation. Analysts take the public verdicts, run a regression, and fill in the missing settlement values to train their models. This approach mathematically guarantees failure. A jury award is almost never the amount actually paid. As David Hyman and Charles Silver documented in their medical malpractice litigation studies, post-verdict settlements, appeals asymmetry, and collectability constraints aggressively compress final paid amounts. If you impute settlements directly from verdicts, you bake survival bias directly into your reserve recommendations. You end up with a model that anticipates a nuclear verdict in every routine slip and fall, driving up unnecessary capital allocation.

This is where the current obsession with large language models creates profound systemic risk for insurance carriers. Generative AI is a token prediction engine. If you prompt a model to read a case file and estimate a settlement value, it generates a number that structurally mimics the text it ingested during its pre-training phase. Because nuclear verdicts, third-party litigation funding, and social inflation generate endless news articles and plaintiff press releases, the model outputs an estimate heavily skewed by that public noise. It does not know the difference between a posturing demand letter and a cleared check. Generation is not prediction. It is an exercise in statistical parroting.

Parsing the Geometry of the Claim

We separate reading and predicting at the fundamental architectural level. We use generative AI strictly as a perception layer. A single claim file contains thousands of pages of pleadings, medical records, deposition transcripts, and opposing counsel correspondence. The large language model reads this corpus and maps the unstructured text into a rigorous taxonomy. It extracts the structured facts, the specific medical codes, the venue characteristics, the procedural posture, and the involved parties. It does the reading, but it absolutely does not do the math. Relying on text generation to calculate a reserve is like asking a dictionary to solve a differential equation.

Once the generative layer structures the case, a separate mathematical machine-learning model takes over. This model is trained exclusively on large numbers of resolved cases with known outcomes. Because we cannot invent missing settlement data where it does not exist, we abandon statistical imputation entirely. We triangulate. Triangulation means defining the boundary conditions of the claim using known structural constraints rather than trying to force a point estimate out of incomplete historical data. We construct a geometric representation of the claim and locate it within a broader mathematical space of historical resolutions.

We map the extracted facts against field-specific priors, liability probabilities, and rigid policy limits. We measure the multidimensional distance between the current case and comparable resolved cases. Crucially, we keep verdict-heavy comparables in a strictly isolated module from settlement-anchored data. Mixing the two contaminates the baseline and destroys the calibration of the model. The dynamics of a jury trial are fundamentally different from the dynamics of a mediation room. Triangulation respects the sparsity of the data by calculating the physics of the negotiation. We assess how similar cases behave under similar constraints rather than blindly applying an arbitrary discount factor to a public jury award.

Honest Uncertainty and Conformal Ranges

A point estimate in a sparse data environment is a mathematical fiction. Claims executives need to allocate defense spend, detect escalation early, and negotiate from data instead of gut instinct. To do that effectively, they need to know the shape of the risk. We output a calibrated settlement range. This conformal range represents the statistical certainty that the true settlement value falls within a specific, mathematically defensible band based on the evidence available at that exact moment in the lifecycle of the claim.

True calibration requires honest error reporting. If the settlement data in a particular jurisdiction for a specific injury type is extremely sparse, the model does not confidently output a tight range. The range widens. The math tells the truth about its own uncertainty. We provide the specific drivers behind the numbers, traceable directly back to the source documents the generative AI read. We calculate the escalation probability and the reserve delta versus the current reserve. You see exactly why the model is uncertain, and you see the comparable cases defining the edges of that uncertainty.

Claims organizations are fighting a compounding war against reserve volatility, social inflation, and third-party litigation funding. Defending the balance sheet requires setting realistic reserves on day one and continuously updating them as new documents arrive. You cannot do that if your predictive systems are hallucinating values derived from plaintiff marketing or applying naive regressions to biased public data. You must build neural-symbolic architectures that respect the fundamental difference between public noise and private resolution. When a claim escalates, the defense counsel and the adjuster need a shared reality. They need to look at the same conformal range and understand exactly which extracted facts are driving the risk upward. A model must earn its trust by showing its work, mapping every prediction back to a specific sentence in a medical report or a specific line in a pleading. You negotiate against the data you have, but you reserve against the math you can prove.