Log in Legacy app
How it works

Six questions in. A defensible number out.

Predict turns a one-minute case description into a confidence-banded settlement value, jurisdiction-tuned and cited to 5–10 comparable verdicts. Here's the version without the math.

Predict your case See it in a demo Free · 60 seconds · no signup to start
  1. 1

    Describe the case

    Six questions: case type, jurisdiction, injury severity, medical specials, property damage, defendant posture. The same facts the senior partner asks for at the intake meeting.

    ≈ 60 seconds · no signup
  2. 2

    We match the cohort

    The model pulls 5–10 comparable verdicts and reported settlements — same case type, same county or state, same severity tier. The cohort is the methodology.

    Stratified jurisdiction folds · 50 states
  3. 3

    You see the band, not just the number

    A median prediction plus a 90% confidence interval. The number is the headline; the band tells you when the cohort is thin and the answer is less certain.

    90–92% median accuracy (MdAPE), temporal held-out test set, by case type →
  4. 4

    Ship the valuation block

    Inside the trial, every prediction exports a demand-letter valuation block — PDF or Word — with the number, band, cohort, and the four facts that moved the prediction.

    Unlocked in the 14-day trial
Inside the product

From six questions to a defensible number — what each screen looks like.

The intake is built around the band — every answer narrows the confidence interval, visibly. When the last question lands, the case-detail view opens to the headline number, the comparable cohort, and the firm's own acceptance threshold. No back-office report. The workspace.

Step 1–3 · Intake
The intake flow with the live preview panel — every answered question narrows the confidence band in real time.
Step 4 · Case detail
Illustrative example. The case-detail view that opens after the prediction lands — number, jurisdiction cohort, acceptance threshold, comparable verdicts, and the demand-letter export, all on one screen.
Illustrative example

What you actually get back.

Illustrative example, built on a real 312-case Harris County cohort you can open in the trial. Numbers are real model output on these inputs, not a named client outcome.

Soft-tissue rear-end, Harris County, Texas. Mid-five-figure med-specials, six months of treatment, moderate severity tier. A carrier first offer of $48,000 on a case like this is the kind of anchor Predict is built to test.

Predict returns the band below in about 30 seconds, sourced to 312 same-jurisdiction same-severity Harris County settlements from the past five years. That cohort is what you cite under the demand-letter valuation block.

  • The number is the headline. $148,000 — the median of what comparable cases settled at.
  • The band is the methodology. ± $22,000 means 9 of 10 similar cases settle in the range.
  • The cohort is the proof. 312 cases, same county, same severity, same case type — openable in the trial.
How the band is computed →
$148,000
± $22,000 · 90% CI
TX · HARRIS · MVA
A $48,000 first offer sits well below this band. The number is the headline; the band is the methodology, and both are cited to the cohort.
What's in the prediction

Three things Predict puts on the plaintiff's side of the table.

Insurers value cases with data and claim history. An attorney working from gut and a verdict search usually does not have the same footing. Every Predict output ships with the same three load-bearing pieces, and each one answers a question an adjuster, a partner, or a judge will eventually ask.

Confidence-banded predictions

The number is the headline. The band is the methodology.

Every prediction comes with a 90% confidence interval — derived from quantile regression against the comparable cohort, not bolted on. The band widens when the cohort is thin or the severity is catastrophic. It's the part of the output that survives the carrier's "where did you get this number?" — and the part that tells you when to slow down before signing the retainer.

9 of 10 cases with the same input profile settle inside the published range. The other 10% are the long tail — flagged, not hidden.

How the band is derived →

Jurisdiction-tuned

Houston isn't Dallas. The model knows.

Jurisdiction is the single largest factor in settlement value — state-level PI economics differ by an order of magnitude, and a jurisdiction-agnostic model would over-predict no-fault states and under-predict high-recovery tort counties. Predict trains with 50 stratified state folds, with additional county-level folds for the 14 highest-volume counties (Harris, Cook, Los Angeles, Bronx, Miami-Dade, and ten more).

50 states · 14 county-level folds · lower-density jurisdictions ship with a wider confidence band by design.

How the folds work →

Cited comparable verdicts

5–10 verdicts behind every number.

Inside the trial, every prediction cites the comparable cohort — 5 to 10 verdicts and reported settlements from the same jurisdiction, same case type, same severity tier. The citation includes outcome amount, jurisdiction, severity, and medical specials. The artifact you paste under the demand-letter valuation block when the adjuster asks for the basis.

312K verdicts in the training set · plaintiff-side only · no carrier data ever.

How the cohort is sourced →

A verdict database hands you a list. Predict hands you a number you can defend.

A verdict-search database returns a stack of cases and leaves the math to you, with no way to tell when the comparable set is too thin to lean on. Predict matches the comparable cohort, returns a single calibrated number with a 90% band, and widens the band, or declines to answer, when the comparable set is too thin to defend. The list is the input. The defensible number is the output.

Built around the moments that actually move the case.

Predict isn't a back-office report. It's an instrument designed for three specific moments where the case value gets decided — or fumbled.

At intake

Run the case in 30 seconds before the retainer goes out. Triage by number, not by gut.

At the demand letter

Anchor the demand to a defensible band cited to comparable verdicts. The carrier responds to methodology.

At partner review

Same number across the firm. The senior reviews methodology, not math.

Quick answers

The short version of the methodology page.

How accurate is the model?

90–92% on the held-out test set, by case type — 92% MdAPE on motor vehicle accidents, 91% on premises liability. The test set is strictly temporal (training data ends 2023-12-31; test set is 2024–2025). Recalibrated quarterly.

What's the training data?

312,000 jury verdicts and reported settlements from 2018–2025, drawn from 50 US jurisdictions. Plaintiff-side and publicly-available records only. No claims-side carrier data is used in training, and none will be.

What case types are supported?

Motor vehicle accidents and premises liability. Medical malpractice, mass tort, workers compensation, and family law are out of scope at launch — and the model won't produce a prediction outside its calibrated case types. We'd rather say "not yet" than ship a number we can't defend.

What about my jurisdiction?

All 50 states are modeled with stratified jurisdiction folds. The 14 highest-volume counties (Harris, Cook, Los Angeles, Bronx, Miami-Dade, Alameda, Fulton, Maricopa, Philadelphia, Dallas, Bexar, Cuyahoga, Kings, Queens) have additional county-level folds. Lower-density jurisdictions ship with a wider confidence band by design.

What if Predict gets a case wrong?

If a prediction lands outside the 90% confidence band, we recalibrate that jurisdiction and disclose it. Brand commitment, not a marketing line. The methodology page documents the recalibration policy and the current model version.

Does Predict replace my judgment?

No. Predict is a statistical estimate of likely settlement value — not a legal opinion, not a trial-verdict prediction, not case strategy. The attorney owns the legal judgment; the model owns the math.

No customer logos here yet, on purpose. Predict launched in 2026, so we don't have a wall of attorney numbers to show. Instead we show the work: the held-out test methodology, the 90% confidence interval math, and the quarterly recalibration log are all published in full on the methodology page. Every number on this page traces back to it.

Read the full methodology →

Pricing

One number. One plan. Every seat included.

A 14-day trial that doesn't ask for a credit card, followed by a flat monthly price that doesn't punish the firm for growing.

$0
First prediction · no signup
14 days
Free trial · no credit card
$499/mo
Unlimited predictions · all seats
See full pricing →