How Predict's case valuations are made.

Training data

The Predict model is trained on 312,000 jury verdicts and reported settlements from 2018 through 2025, drawn from 50 US jurisdictions. The dataset is plaintiff-only — every record is sourced from publicly available verdict and settlement filings, plaintiff-side reported settlements, and PACER docket extracts. No claims-side carrier data is used in training, and none will be.

The 312K count is the raw record count before deduplication. After deduplication and the 4-year rolling-window training cut-off, the active training set is roughly 248K records, of which 198K are motor vehicle accident cases and 72K are premises liability cases. The remaining 50K records are out-of-scope case types (medical malpractice, mass tort, workers compensation) carried for transfer-learning signal but not used in production predictions.

Per-record attributes: case type, jurisdiction (state and county), liability classification, injury severity tier, medical specials, property damage, plaintiff-counsel firm, judge composition, defendant insurance posture, and outcome amount with verdict-versus-settlement classification.

Model architecture

The production model is a gradient-boosted regression ensemble — 800-tree XGBoost trained with monotonicity constraints on the severity and medical-specials features. Hyperparameters were tuned via 5-fold cross-validation on the pre-2024 training portion, with the 2024–2025 records held out as the test set.

Gradient-boosted regression was selected over neural network alternatives for two reasons. First, the feature space is tabular and modestly sized — the strong inductive bias of trees on tabular data is well-established. Second, interpretability matters in legal contexts more than in most ML applications: an attorney must be able to defend the number, which means understanding the feature contributions. The model produces per-prediction SHAP attributions inside the full Predict subscription, surfaced as "the four facts that moved the number" alongside each prediction.

Accuracy and held-out testing

The accuracy figure cited across the site — 90–92% — is the Median Absolute Percentage Error (MdAPE) on the held-out test set, by case type:

MdAPE is the metric most appropriate for skewed long-tail damages distributions — mean-based error metrics over-weight the catastrophic tail and under-represent typical case performance. The methodology mirrors the held-out test design used in the underwriting-model literature.

The held-out split is strictly temporal — training data ends 2023-12-31, the test set is 2024-01-01 through 2025-12-31. No leak between train and test.

What this means at your desk: when a case sits well below the bottom of its 90% band, that is the model telling you a number deserves a second look before you accept it. The headline is the number; the band is where the case should land 9 times out of 10 on the same facts.

Confidence-band derivation

Every Predict prediction ships with a 90% confidence interval. The band is computed from quantile regression — a parallel set of 800-tree XGBoost models trained against the 5th and 95th percentiles of the conditional outcome distribution, rather than the median.

The band widens with two conditions: (1) injury severity — catastrophic cases carry materially wider conditional distributions than minor cases; (2) data density — predictions in jurisdictions with fewer than 100 comparable verdicts in the training set widen by a density correction factor.

Operational interpretation: 9 out of 10 cases with the same input profile settle inside the published range. The remaining 10% are the long tail — high-stakes outliers, judges with unusual records, plaintiff counsel with extraordinary individual histories.

Want to see this on one of your own cases? Start a 14-day trial, no credit card, or try the public calculator first.

Jurisdiction folds

Jurisdiction is the single largest factor in settlement value. The model treats jurisdiction with stratified folds — 50 state-level folds, with additional county-level folds for the 14 counties with ≥150 verdicts in the training set (Harris TX, Cook IL, Los Angeles CA, Bronx NY, Miami-Dade FL, Alameda CA, Fulton GA, Maricopa AZ, Philadelphia PA, Dallas TX, Bexar TX, Cuyahoga OH, Kings NY, Queens NY).

The fold architecture matters because state-level PI economics differ by an order of magnitude. A jurisdiction-agnostic model would over-predict no-fault states and under-predict high-recovery tort jurisdictions. The stratified fold approach is what allows the 90–92% MdAPE figure to hold across the 50-state range rather than collapsing to an aggregate average.

Recalibration policy

If a prediction misses outside the published 90% confidence band, the model is recalibrated for that jurisdiction and the recalibration is disclosed. This is a brand commitment, not a marketing line.

Recalibration cadence: continuous on the verdict-feedback side (new outcomes are scored against prior predictions weekly), with a full model refresh on a quarterly cadence. Each quarterly refresh ships with a changelog documenting the per-jurisdiction MdAPE delta versus the prior model.

The model versioning is public. The current production model is v3.2, refreshed 2026-04-01. The next scheduled refresh is 2026-07-01.

Sourcing and citations

Every prediction inside the full Predict subscription cites 5–10 comparable verdicts drawn from the training set. The citation includes the jurisdiction, case-type classification, severity tier, medical specials, and outcome amount — anonymized to the case-caption level but specific enough to support the prediction defensively.

The annual State of PI Case Valuation report sources every chart and every claim against the underlying training data. Every accuracy claim names the held-out test set. Every jurisdiction stat names the year and the N.

What the model does not do

• The model is not a legal opinion. Predictions are statistical estimates of likely settlement value, not legal advice on case viability, recovery strategy, or settlement positioning. The attorney owns the legal judgment.
• The model is not calibrated outside MVA and PL. Medical malpractice, mass tort, workers compensation, family law, and immigration cases are out of scope at launch. Predictions on out-of-scope case types are not produced.
• The model does not predict trial outcomes. The output is the expected settlement value, not the expected jury verdict. The two diverge materially in many jurisdictions — and where the divergence is large, the band widens to reflect it.
• The model does not know about cases that haven't been filed. Predictions are calibrated against the historical training set; emerging case types (autonomous-vehicle cases, new statutory tort regimes) are added to the training set with each quarterly refresh, with appropriately wider bands until the data density supports tightening.