The Predictable Court

Abstract

To date, the best-performing legal prediction technology remains the oldest: the human mind. Human crowds consistently outperform algorithmic approaches to forecasting judicial outcomes. This Article challenges that arrangement. Using a fine-tuned transformer architecture trained on fifteen years of unstructured Supreme Court data, the Article introduces the “Experience Model,” which learns the personalities, preferences, and jurisprudential commitments of the sitting justices. Tested against the most recent term (OT2024), the model achieves 74 percent vote-level accuracy and 82 percent case-level accuracy—outperforming existing benchmarks and, at the case level, human crowds in the test term. Beyond mere forecasting, the model’s calibrated predictions point to a novel methodology for addressing longstanding questions of legal theory and institutions, illustrated in four applications. First, legal theorists from Hart to Raz prize law’s guidance function—or capacity to provide citizens a discernible pathway for their conduct—a concept tied to predictability. By this measure, law’s guidance function operates more fully in some domains than others: for instance, criminal procedure cases are difficult to predict; judicial power cases are highly predictable. Second, the model provides a novel methodology for testing the Priest- Klein hypothesis that litigated cases tend to be uncertain propositions—using more theory aligned but previously unobservable calibrated ex-ante win probabilities rather than observed win rates. Results largely support the hypothesis, as adjusted for docket selection through the certiorari process. Third, in the first justice-level study of the marginal predictive contribution of oral argument, an ablation study demonstrates that oral argument substantially boosts predictive accuracy, but does so heterogeneously, with large gains for some justices and negligible effects for others. Fourth, cases decided in June, often considered the term’s most complex, turn out to be its most predictable. This suggests that, though controversial, end-of-term cases divide along ideological lines and other contours the model easily tracks rather than presenting truly uncertain legal or coalitional questions. The Article concludes by examining the normative implications of accurate prediction. Some predictive capacity enhances law’s guidance function, but over-reliance risks reducing innovation in legal doctrine and displacing authoritative judicial reasoning. Legal institutions,however, are not passive subjects of technological disruption. Courts control the data on which prediction models depend—including panel disclosure timing, authorship norms, and decision structure—and can titrate that information flow to foster a healthy legal system.