Conformal Prediction for Claims: Ranges, Not Point Guesses

TL;DR — Point predictions fail in claims because they ignore litigation variance. Conformal prediction solves this by outputting a calibrated range with a mathematical guarantee of accuracy, allowing executives to set realistic reserves without false certainty.

If an AI vendor tells you their model predicts a bodily injury claim will settle for exactly $435,000, they are either mathematically illiterate or lying. Litigation is fundamentally probabilistic. There is no deterministic physics equation governing a slip-and-fall claim or a commercial auto accident. Human judges, aggressive plaintiff attorneys, and shifting jury sentiments dictate final outcomes. Providing a single point guess in this high-variance environment gives the illusion of precision while guaranteeing error. You will be wrong. The only question is by how much. In an era defined by social inflation and third-party litigation funding, being wrong by a small margin is rare. Being wrong by millions is systemic.

Claims organizations rely on accurate reserves to survive. When actuaries and claims leaders build those reserves around static point estimates, they bake fragility directly into the balance sheet. A single number cannot express variance. If an adjuster anchors on a $150,000 prediction and the plaintiff firm escalates the tactics, the defense strategy collapses. Capital is either locked up unnecessarily or disastrously under-allocated. We require a different mathematical framework for risk. We must stop predicting points and start predicting regions. The goal of machine learning in insurance is not to guess a number. The goal is to accurately quantify the uncertainty.

The Mathematics of Honest Doubt

In statistics, conformal prediction is a framework that turns any heuristic model into a rigorously calibrated predictor. Instead of outputting a single target number, a conformal predictor outputs a set or a range. It provides a mathematical guarantee of coverage. If you ask the system for a 90 percent confidence interval, the true outcome will fall inside the predicted range exactly 90 percent of the time, assuming the underlying data distribution remains stable. It achieves this by calculating non-conformity scores. The model examines its past errors on similar data and explicitly widens or narrows its output range based on how strange the new data point looks compared to the training set. It is a system designed to report honest error.

At Canotera, we enforce a strict architectural boundary between reading text and predicting financial outcomes. Generative AI is excellent at parsing a thousand-page medical file or extracting the severity of a spinal injury from messy litigation pleadings. It does the reading and the structuring. It is fundamentally incapable of actuarial prediction. Large language models hallucinate numbers because they predict the next word in a sequence, not the actual financial exposure of a casualty claim. We use generative models exclusively for extraction. We map the unstructured text from pleadings, medicals, and correspondence into a deterministic, neural-symbolic geometry.

Once the generative model structures the case file, separate machine learning models take over. These geometric models are trained on large numbers of resolved cases where the final outcomes are known realities. By measuring how similar the new claim is to historical cases in this structured space, the system calculates a baseline. Instead of stopping at a naive point estimate, we apply conformal prediction to generate a calibrated settlement range. The system calculates the margin of error based on the density and variance of the comparable resolved cases. If a claim sits in a sparse area of the historical data, the range widens to reflect the lack of precedent.

Calibration and Claims Reality

A model is calibrated if its predicted probabilities match observed real-world frequencies. If our platform states that a specific class of claims has a 70 percent chance of escalation to litigation, exactly seven out of ten such claims should escalate. Calibration allows a claims executive to trust the output. When a new file hits the desk and the platform outputs a settlement range of $200,000 to $350,000, you know the boundaries are mathematically sound. You can set realistic reserves on day one. You can identify the reserve delta against your current estimates immediately, preventing surprises before trial.

This approach changes how defense spend is allocated. Narrow ranges indicate high confidence and dense historical precedent. The adjuster knows they can move quickly to settle, avoiding protracted legal fees. Wide ranges indicate high uncertainty. This variance is usually driven by missing medical documentation, an unknown plaintiff firm, or compounding severity factors. The platform surfaces the specific drivers behind the uncertainty. The adjuster knows exactly why the range is wide and what specific information is required to narrow it. You negotiate from a position of structural data rather than gut instinct.

Every number in the platform must be traceable to the source documents. A settlement range is useless if the claims handler cannot see the underlying math and the comparable evidence. Because the prediction models are completely distinct from the language models, we can surface the exact resolved cases that anchor the range. The adjuster sees the historical geometry. They see the specific medical codes, jurisdiction factors, and plaintiff attorney behaviors driving the calculation. The AI does not ask for blind faith. It presents a statistical argument backed by known outcomes.

The insurance industry is fighting rising severity, nuclear verdicts, and systemic reserve volatility. Relying on fragile point estimates in this environment is a mathematical failure. True machine learning does not attempt to eliminate uncertainty through false precision. It measures the uncertainty, bounds it, and makes it actionable for the business. A model that admits what it does not know is the only one you can trust with your balance sheet. Stop asking algorithms to predict the future, and start asking them to map the boundaries of your exposure.