Building Forecasts When Comparables Report Verdicts

TL;DR — Public verdicts represent extreme anomalies, not standard settlement values. Accurate claims forecasting requires decoupling language models that read case files from the mathematical models that calculate settlement ranges based on resolved outcomes.

The fundamental flaw in legal analytics is treating a verdict like a price tag. A hundred-million-dollar jury award makes headlines, but it rarely dictates the wire transfer that closes the file. If you build a forecasting system that feeds raw verdict data directly into a settlement prediction model, you guarantee systematic over-reserving. Claims professionals know the public record stops at the gavel. It misses the high-low agreements, the post-trial negotiations, the appeals asymmetry, and the hard ceiling of policy limits. When engineering a predictive platform, treating these public verdicts as ground-truth settlement comparables introduces fatal skew into the entire data model.

The cases that go to trial suffer from severe selection bias. Under the Priest-Klein hypothesis, disputes that reach a jury are the anomalies where both sides hold mutually exclusive, highly confident views of the outcome. They are the hardest files to settle, often involving complex liability disputes, untested legal theories, or extreme damages. Building a forecasting model on this skewed dataset is like estimating average daily weather by only measuring hurricanes. The system must account for the fact that a jury award is almost never the amount paid. It is a theoretical maximum used as a weapon in negotiation, not a baseline for standard file resolution.

Isolating the Signal from the Noise

Solving this data problem requires enforcing a strict architectural boundary between reading data and predicting outcomes. Generative AI is highly effective at parsing unstructured text, but it is terrible at math. At Canotera, we point our language models at the raw case files. The system processes pleadings, medical chronologies, and correspondence to extract structural facts. It identifies injury severity, jurisdiction, plaintiff characteristics, and specific liability drivers across thousands of pages of raw text. The generative AI does the reading and the structuring. It does not predict the outcome. Language models are autocomplete engines, highly susceptible to the availability cascade of publicized nuclear verdicts. We keep them entirely out of the prediction layer.

Prediction happens in a separate mathematical machine-learning layer. This model maps the extracted facts to a realistic settlement range. The primary engineering challenge here is data sparsity. Confidential settlements outnumber public verdicts by a massive margin, but that data remains locked in private carrier vaults. To build a reliable forecast without relying on skewed public data, we triangulate. We rely on field-specific priors and honest uncertainty bands. We train our geometric models on large numbers of resolved cases where the actual paid outcome is known. This physically decouples the public verdict headlines from the private settlement reality, ensuring the math remains grounded in actual risk transfer mechanics.

Structuring the Comparables

When an adjuster evaluates a file, they need comparables to justify their reserve delta and negotiate effectively. Surfacing a list of nuclear verdicts as direct comparables destroys the adjuster's credibility in a negotiation. Plaintiff counsel will immediately recognize the numbers as outliers. We handle this by keeping verdict-heavy comparables in a completely separate module from settlement-anchored data. The user sees a calibrated settlement range, but they also see exactly which historical cases drove that calculation. Every comparable is traceable back to the source documents. If a specific medical code or liability trigger shifts the forecast, the system highlights the exact paragraph in the uploaded file.

A pragmatic system must also account for the physical limits of the real world. A geometric model predicting a massive exposure fails in practice if the policy limit is capped and the defendant lacks liquid assets. The mathematical layer incorporates collectability constraints directly into the calculation. It models the probability of liability separately from the severity of damages. By breaking the forecast into discrete, accountable components, we give the claims team a tool they can defend in a committee review. This granularity allows carriers to allocate defense spend accurately, fighting the claims that require litigation and settling the ones that do not.

Designing for Realistic Reserves

The objective of this architecture is setting an accurate reserve on day one. Volatility in reserving destroys capital efficiency across the entire insurance enterprise. Over-reserving based on the fear of social inflation and third-party litigation funding ties up cash that should be invested. Under-reserving leads to massive adverse development later in the lifecycle of the claim, shocking the balance sheet. Our platform delivers a settlement range rather than a single point guess to reflect the inherent uncertainty of litigation. A narrow band indicates high confidence and dense settlement data. A wide band signals escalation risk, requiring immediate defense spend allocation to investigate the unknowns and close the information gap.

Building a reliable forecasting engine requires ignoring the noise surrounding artificial intelligence. You cannot feed a stack of PDFs into a neural network and expect a realistic financial forecast. You have to build a deterministic data pipeline. You parse the facts, isolate the selection bias, constrain the math with actual settlement physics, and expose the reasoning to the user. The platform must provide the specific drivers behind every number so the claims team can negotiate from data instead of gut instinct. Transparency is a security requirement for sensitive records, and traceability by design is the only way to build trust with the professionals handling the files.

A prediction is only useful if you can trace the math backward to a fact that actually happened.