Why We Output Bands When the Literature Can't Give You a Point

TL;DR — Single-number predictions hide the volatility of litigation. By outputting calibrated settlement ranges, claims teams can mathematically distinguish between highly predictable files and volatile cases requiring immediate defense spend.

The most frequent request we receive from a claims desk is a demand for a single dollar figure. Adjusters and executives want the exact settlement value for a newly opened file. The mathematics of litigation refuse to provide it.

A point estimate in litigation forecasting is a fiction. When a software tool outputs exactly $432,500, it is lying about its own certainty. The legal system is inherently stochastic. Liability probability is rarely binary. Plaintiff credibility is highly subjective. The increasing presence of third-party litigation funding removes the plaintiff's liquidity constraints, altering the settlement threshold and widening the variance of possible outcomes. Two distinct juries, or two mediators operating on consecutive days, will yield wildly different financial outcomes for the exact same set of facts. Providing a single number strips away the variance that defines the risk.

The generative hallucination

Current artificial intelligence trends do immediate damage to this problem. Large language models are text generators. They predict the next token in a sequence based on training distributions. If you feed thousands of pages of pleadings, medical records, and correspondence into an LLM and prompt it for a case value, it will confidently print a number. It is not calculating probability. It is pattern-matching legal syntax to generate a dollar amount that looks structurally plausible. A language model lacks an internal world model of financial liability. It cannot weigh the historical probability of a cervical fusion in Cook County against a similar fusion in a conservative rural venue. It merely generates text that sounds like a claims adjustment.

We strictly separate the act of reading from the act of predicting. Generative AI excels at reading. We use it to parse the unstructured mess of a case file, extracting specific injury types, venues, liability indicators, and demographic variables. Once the data is structured into a symbolic representation, the language model stops. The actual prediction relies on geometric machine learning models trained exclusively on large numbers of resolved cases with known outcomes. This neural-symbolic structure ensures that the reading mechanism cannot hallucinate the math. The geometric models operate purely on mathematical distances between the structural features of the new case and the historical baseline.

The selection bias of the public record

Even with a dedicated mathematical model, predicting a settlement is notoriously difficult because of the underlying data landscape. The Priest-Klein hypothesis demonstrates that cases reaching a trial verdict are a non-random, heavily biased sample. They represent the specific subset of disputes where the plaintiff and defense fundamentally misaligned on liability or damages. Verdict data is highly visible and public, but relying on it to predict early-stage settlements anchors your model to the outliers.

Most cases settle. Yet settlement data is private, sparse, and noisy. Adding to the friction, a jury award is rarely the amount actually paid. The empirical work of researchers like David Hyman shows that appeals asymmetry, policy limits, and collectability force post-verdict settlements that look nothing like the headline nuclear verdict. A mathematical model must navigate this difference. It requires keeping verdict-heavy comparables in a strictly separate module from settlement-anchored data. If you mix them, the long tail of nuclear verdicts pulls the mean upward, destroying the model's accuracy on standard claims.

Optimizing a model for a single point estimate on a skewed distribution guarantees failure. If you train a neural network using mean squared error on litigation data, the prediction gets dragged toward the extreme outliers. The model learns to over-predict standard cases to avoid massive penalties on the rare nuclear outcomes. The result is a point estimate that is mathematically wrong for both scenarios.

Honest error reporting

Because the underlying system is volatile, we output a calibrated settlement range. We rely on conformal prediction techniques to establish honest uncertainty bands. Conformal prediction works by analyzing the heuristic error of the model on a calibration dataset. It then outputs a set of values rather than a single point, mathematically guaranteeing that the true settlement value will fall within the predicted band at a specified confidence level. This is not a point guess with an arbitrary percentage tacked on as a margin of error. It is a rigorous quantification of doubt. A model is calibrated if its 80 percent confidence interval actually contains the true outcome 80 percent of the time. Alongside the range, we output an escalation probability, comparable resolved cases, and a reserve delta against the current baseline, calculating the specific drivers behind each number traceable directly to the source documents.

A calibrated range is entirely actionable. A tight band tells you the model has seen this specific geometry of facts thousands of times. You can confidently set your day-one reserve and initiate data-driven settlement talks. A wide band is an alarm bell. It signals high volatility and low certainty. It tells the claims executive to allocate defense spend immediately to close the information gap, investigate liability harder, or prepare for the specific risk of social inflation driving an extreme outcome.

You cannot negotiate effectively if you are blind to the shape of the risk. A single number hides the volatility. A conformal range exposes it. We refuse to output a point estimate because the goal of machine learning in insurance is not to guess a number. The goal is to map the probability space.