Outcome forecasting
We design prediction tasks around disputes where the decision path can be inspected, not just guessed from labels.
AI Systems That Forecast Legal Outcomes
Research notes and challenge reports
Legal outcome prediction, tested against real disputes
CaseCrunch Lab studies whether automated reasoning systems can predict legal outcomes with enough discipline to assist lawyers, researchers, policymakers, and justice technologists.
The work starts with a practical question: can a machine read the same dispute material as a trained lawyer and reach a defensible forecast before the final decision is known? That question sounds narrow, but it touches almost every hard problem in legal AI: factual ambiguity, incomplete records, shifting standards of review, and the difference between a plausible argument and a likely result.
Automated reasoning
The lab studies how systems represent facts, claims, legal rules, and uncertainty when a case does not fit a clean template.
Human comparison
Controlled lawyer-versus-system exercises help separate model confidence from useful legal judgment.
How the research is run
One completed prediction exercise may begin with fewer than a few hundred disputes, but the work around those files is exacting. Each matter needs a stable record, a defined question, a time cut-off, and an outcome that can be checked after prediction.
From case record to prediction task
According to measurements from completed exercises, the strongest signal rarely comes from a single dramatic fact. It comes from the structure of the claim: who bears the burden, what evidence was available, which remedy was requested, and whether the decision-maker had followed similar reasoning before.
Dataset design is therefore treated as legal work, not clerical preparation. A case file becomes useful only when the relevant facts, procedural posture, decision horizon, and scoring rule have been fixed before the prediction is made.
Important:
Prediction quality still turns on the framing of the dispute and the completeness of the record. A model can score a defined legal question; it cannot repair a badly specified one.
Calibration and scoring
Calibration confirmed whether confidence tracked reality, not whether a model sounded persuasive. In controlled trials, this distinction mattered: a system that gave modest probabilities for difficult cases often served the research question better than one that gave bold answers across the board.
The technical side of this work connects closely with evaluation metrics for legal reasoning engines, where accuracy, calibration, coverage, and error type each tell a different part of the story.
Research areas across legal AI
Some projects ask whether a system can beat a human prediction panel. Others ask a quieter question: which parts of legal knowledge can be represented in a machine-readable form without stripping away the context that makes the dispute legally meaningful?
Research governance
The first approach works well for public demonstrations and tightly scoped contests. The second is slower, but it gives better material for court administrators, regulators, and researchers who need to understand why a system reached a forecast. CaseCrunch Lab uses both approaches where they fit.
Prediction Challenges
AI versus human lawyer contests and outcome forecasting experiments.
Automated Reasoning
Machine-driven legal analysis, rule representation, and decision modeling.
Consumer Claims
Financial mis-selling, PPI, and ombudsman-style dispute prediction.
Justice Automation
The implications of AI for legal knowledge and access to justice.
Research Methodology
Datasets, experimental design, scoring frameworks, and evaluation practice.
Evidence before claims
Legal AI attracts confident language quickly. CaseCrunch Lab takes a narrower route: define the task, run the comparison, publish enough method for scrutiny, then explain what the result can and cannot support.
Recorded results show why this discipline matters. A model may perform well on consumer redress disputes and still need different testing before anyone relies on it for judicial triage, settlement valuation, or regulatory review.
Field Note:
- Cambridge-origin research programme focused on automated reasoning and legal outcome prediction.
- Scoped work across prediction challenges, consumer claims, financial mis-selling, and ombudsman-style disputes.
- Method-led publication practice for legal professionals, AI researchers, policymakers, and justice technologists.
The lab director and research contributors bring expertise across legal analysis, machine learning evaluation, dispute datasets, and justice-system design. Their shared concern is not whether automation can replace legal judgment. It is where machine prediction can make human judgment more consistent, auditable, and empirically grounded.
Bottom Line:
Legal outcome prediction is useful when the question is precise, the dataset is inspectable, and the system is tested against the decision standard it claims to forecast.
Our Research Process
Research
Gather and review existing literature and data.
Analyze
Apply rigorous methods to evaluate findings.
Publish
Share results with the scientific community.