The Supreme Court — the record
The model's full walk-forward backtest, 1970–2023 (53,291 justice-votes), and how it stacks up against the published academic benchmark and every other forecaster we track. ← Back to the current court
How well it forecasts
Scored strictly walk-forward over 1970–2023. The honest bar is the Court's strong habit of siding with the petitioner (it reverses more than it affirms) — so "petitioner always wins" is a tough baseline, and it has only gotten tougher as recent benches grew more lopsided. The durable edge is at the justice level.
Per-justice vote accuracy vs the baseline, by term
The model holds a gap above the "petitioner always wins" line across the era. The gap narrows recently because the baseline itself climbs — modern terms break the petitioner's way more often, leaving less room to beat.
| By issue area | Model | Base | n |
|---|---|---|---|
| Criminal Procedure | 67.8% | 68.7% | 1,471 |
| Economic Activity | 63.6% | 63.4% | 1,087 |
| Civil Rights | 63.7% | 64.2% | 1,084 |
| Judicial Power | 63.3% | 61.6% | 804 |
| First Amendment | 67.2% | 67.0% | 436 |
| Federalism | 61.3% | 64.1% | 287 |
| Due Process | 69.0% | 69.4% | 271 |
| Unions | 57.7% | 59.3% | 194 |
| Federal Taxation | 64.8% | 55.2% | 145 |
| Privacy | 69.0% | 69.0% | 116 |
| Attorneys | 67.8% | 65.5% | 87 |
| Interstate Relations | 55.4% | 51.4% | 74 |
All models, compared
Every rule that turns information into a prediction is a model. That includes VotePredictor, the FantasySCOTUS crowd, the published {Marshall}+ method, and the simple "petitioner always wins" model. We group all of them here. Scores are only ranked within identical evaluation samples; different periods, cases, or prediction units are shown together but not treated as a head-to-head race.
Long-run models · same 5,914 cases · 1971–2023
VotePredictor and our faithful reproduction of {Marshall}+ are trained walk-forward. The petitioner-always-wins model is a deterministic base-rate model. All three are scored on the same underlying cases below.
VotePredictor SCOTUS vs {Marshall}+ (reproduction)
same cases · 1971–2023 · 5,914 cases| Metric | VotePredictor SCOTUS | {Marshall}+ | Petitioner-always-wins model |
|---|---|---|---|
| Justice-vote accuracy | 67.2% | 64.9% | 60.6% |
| Justice-vote Brier (lower better) | 0.208 | 0.218 | — |
| Case accuracy (vote aggregation) | 67.3% | 65.6% | — |
| Justice-vote accuracy by decade | 1970s | 1980s | 1990s | 2000s | 2010s | 2020s |
|---|---|---|---|---|---|---|
| VotePredictor SCOTUS | 68.0% | 67.4% | 65.0% | 67.6% | 66.9% | 69.1% |
| {Marshall}+ | 66.0% | 66.6% | 61.9% | 64.4% | 62.8% | 67.7% |
| Petitioner-always-wins model | 61.4% | 58.0% | 57.2% | 64.2% | 63.4% | 65.5% |
{Marshall}+is our faithful reproduction of the Katz–Bommarito–Blackman (2017) method — extremely-randomized trees on the full set of raw SCDB codes — since their actual per-case predictions aren't published. Their reported headline was 71.9% justice / 70.2% case over 1816–2015; that span has a lower base rate than the modern window here, so it's context rather than a like-for-like target. On identical modern cases, our model edges the classic approach.
Crowd model · recovered same-case sample · 2019–2024
FantasySCOTUS is a crowd-aggregation model, not a separate category. We recovered its public pages from the Internet Archive and kept only snapshots captured before the decision. The cleanest comparison has 39 direct case-outcome crowds from October Terms 2019–2024. Archive coverage is incomplete, so this is an audited sample, not directly comparable to the long-run table above.
| Same archived cases | n | accuracy | Brier |
|---|---|---|---|
| FantasySCOTUS crowd model | 39 | 69.2% | 0.225 |
| VotePredictor SCOTUS | 39 | 66.7% | 0.238 |
A strictly forward blend—its crowd weight chosen using earlier terms only—scores 72.0% accuracy and 0.222 Brier on 25 later cases, versus 68.0% / 0.234for the crowd and 68.0% / 0.232for our model on those same rows. That result is promising but small-sample. Another 53 cases have only per-justice crowd probabilities; they remain labeled as a separate reconstructed series rather than being passed off as direct case forecasts.
The model zoo — every forecaster, ranked
Rather than pick one design, we build them all and let the data rank them — forecasting October Term 2025 cases after oral argument. Our model is VotePredictor SCOTUS (a per-justice model on the Supreme Court Database — separate from the VotePredictor Elections model), in a structural variant and one that adds oral-argument analysis, alongside the FantasySCOTUS crowd, prediction markets, and a consensus blend. Two honest takeaways: the informed crowd wins, and analyzing the oral argument is the strongest single signal we add — historically beating our structural model on its own.
| Forecaster | Accuracy | Brier | n |
|---|---|---|---|
| FantasySCOTUS crowd | 95.8% | 0.124 | 24 |
| Consensus (model + crowd + market)not independent | 91.7% | 0.142 | 24 |
| VotePredictor SCOTUS | 83.3% | 0.189 | 24 |
| VotePredictor SCOTUS + oral argument | 75.0% | 0.201 | 24 |
| "Petitioner always wins" baseline | 83.3% | — | — |
Does analyzing the oral argument help? (2013–2023, walk-forward)
On 640 argued cases with transcripts: the oral-argument signal alone (how hard the justices grill each side) is the strongest single input — it beats our structural model on both accuracy and calibration. None beats the lopsided base rate on raw accuracy, but the signal is real.
| Model | Accuracy | Brier |
|---|---|---|
| VotePredictor SCOTUS (structural) | 60.5% | 0.246 |
| Oral argument only | 63.4% | 0.240 |
| VotePredictor SCOTUS + oral | 65.5% | 0.235 |
| baseline | 67.7% | — |
| Case | VotePredictor SCOTUS | +oral | crowd | market | consensus | result |
|---|---|---|---|---|---|---|
| learning resources v. trump Roberts 63Thomas 60Alito 67Sotomayor 47Kagan 51Gorsuch 65Kavanaugh 65Barrett 62Jackson 55 | 59% | 59% | 57% | — | 58% | Petitioner |
| robinson v. callais Roberts 61Thomas 45Alito 50Sotomayor 76Kagan 76Gorsuch 67Kavanaugh 63Barrett 65Jackson 72 | 64% | 64% | 41% | — | 49% | Respondent |
| chiles v. salazar Roberts 75Thomas 76Alito 77Sotomayor 41Kagan 50Gorsuch 73Kavanaugh 73Barrett 71Jackson 54 | 66% | 61% | 73% | — | 69% | Petitioner |
| wolford v. lopez Roberts 61Thomas 45Alito 50Sotomayor 76Kagan 76Gorsuch 67Kavanaugh 63Barrett 65Jackson 72 | 64% | 65% | 64% | — | 64% | Petitioner |
| noem v. al otro lado Roberts 70Thomas 78Alito 75Sotomayor 38Kagan 46Gorsuch 69Kavanaugh 67Barrett 70Jackson 54 | 63% | 73% | 68% | — | 70% | Petitioner |
| cox communications v. sony music Roberts 55Thomas 52Alito 54Sotomayor 68Kagan 69Gorsuch 60Kavanaugh 58Barrett 59Jackson 66 | 60% | 78% | 94% | — | 88% | Petitioner |
| fcc v. at&t Roberts 55Thomas 52Alito 54Sotomayor 68Kagan 69Gorsuch 60Kavanaugh 58Barrett 59Jackson 66 | 60% | 60% | 70% | — | 67% | Petitioner |
| chevron usa v. plaquemines parish Roberts 73Thomas 67Alito 74Sotomayor 48Kagan 54Gorsuch 72Kavanaugh 78Barrett 74Jackson 64 | 67% | 60% | 72% | — | 68% | Petitioner |
| monsanto v. durnell Roberts 63Thomas 60Alito 67Sotomayor 47Kagan 51Gorsuch 65Kavanaugh 65Barrett 62Jackson 55 | 59% | 69% | 52% | — | 58% | Petitioner |
| hikma pharmaceuticals v. amarin Roberts 55Thomas 52Alito 54Sotomayor 68Kagan 69Gorsuch 60Kavanaugh 58Barrett 59Jackson 66 | 60% | 52% | 93% | — | 78% | Petitioner |
| exxon mobil v. corporacion cimex Roberts 48Thomas 42Alito 39Sotomayor 50Kagan 57Gorsuch 53Kavanaugh 59Barrett 54Jackson 51 | 50% | 75% | 61% | — | 66% | Petitioner |
| bost v. illinois state board of elections Roberts 48Thomas 42Alito 39Sotomayor 50Kagan 57Gorsuch 53Kavanaugh 59Barrett 54Jackson 51 | 50% | 78% | 79% | — | 79% | Petitioner |
| villarreal v. texas Roberts 50Thomas 40Alito 37Sotomayor 71Kagan 68Gorsuch 54Kavanaugh 56Barrett 53Jackson 66 | 55% | 82% | 16% | — | 42% | Respondent |
| barrett v. united states Roberts 50Thomas 40Alito 37Sotomayor 71Kagan 68Gorsuch 54Kavanaugh 56Barrett 53Jackson 66 | 55% | 68% | 76% | — | 73% | Petitioner |
| bowe v. united states Roberts 50Thomas 40Alito 37Sotomayor 71Kagan 68Gorsuch 54Kavanaugh 56Barrett 53Jackson 66 | 55% | 65% | 61% | — | 62% | Petitioner |
| usps v. konan Roberts 73Thomas 67Alito 74Sotomayor 48Kagan 54Gorsuch 72Kavanaugh 78Barrett 74Jackson 64 | 67% | 54% | 55% | — | 55% | Petitioner |
| olivier v. city of brandon Roberts 74Thomas 52Alito 62Sotomayor 76Kagan 78Gorsuch 77Kavanaugh 75Barrett 74Jackson 74 | 71% | 80% | 88% | — | 85% | Petitioner |
| first choice women's resource centers v. platkin Roberts 74Thomas 52Alito 62Sotomayor 76Kagan 78Gorsuch 77Kavanaugh 75Barrett 74Jackson 74 | 71% | 78% | 94% | — | 88% | Petitioner |
| hencely v. fluor Roberts 55Thomas 52Alito 54Sotomayor 68Kagan 69Gorsuch 60Kavanaugh 58Barrett 59Jackson 66 | 60% | 49% | 64% | — | 59% | Petitioner |
| landor v. louisiana dept. of corrections Roberts 74Thomas 52Alito 62Sotomayor 76Kagan 78Gorsuch 77Kavanaugh 75Barrett 74Jackson 74 | 71% | 88% | 48% | — | 63% | Respondent |
| galette v. new jersey transit Roberts 63Thomas 57Alito 56Sotomayor 70Kagan 72Gorsuch 67Kavanaugh 66Barrett 62Jackson 69 | 65% | 49% | 51% | — | 50% | Petitioner |
| t.m. v. university of maryland medical system Roberts 48Thomas 42Alito 39Sotomayor 50Kagan 57Gorsuch 53Kavanaugh 59Barrett 54Jackson 51 | 50% | 76% | 65% | — | 69% | Respondent |
| cisco systems v. doe i Roberts 70Thomas 78Alito 75Sotomayor 38Kagan 46Gorsuch 69Kavanaugh 67Barrett 70Jackson 54 | 63% | 90% | 74% | — | 79% | Petitioner |
| havana docks v. royal caribbean Roberts 55Thomas 52Alito 54Sotomayor 68Kagan 69Gorsuch 60Kavanaugh 58Barrett 59Jackson 66 | 60% | 87% | 60% | — | 69% | Petitioner |
| trump v. slaughter Roberts 79Thomas 76Alito 79Sotomayor 46Kagan 53Gorsuch 79Kavanaugh 74Barrett 79Jackson 56 | 69% | 62% | 63% | 89% | 72% | pending |
| west virginia v. b.p.j. Roberts 75Thomas 85Alito 85Sotomayor 32Kagan 33Gorsuch 57Kavanaugh 59Barrett 55Jackson 44 | 59% | 32% | 71% | — | 58% | pending |
| little v. hecox Roberts 72Thomas 81Alito 78Sotomayor 36Kagan 34Gorsuch 72Kavanaugh 71Barrett 73Jackson 43 | 62% | 90% | 68% | — | 76% | pending |
Cells show each forecaster's P(petitioner wins); green/red= right/wrong on decided cases. Under each case are the model's per-justice predictions — each justice's % chance of siding with the petitioner (name colored by appointing party, D/R; hover for detail). Forecasts are made after oral argument only. VotePredictor SCOTUS is our per-justice model on the Supreme Court Database — a separate model from VotePredictor Elections. Consensus blends model + crowd + market, weighted by each source's track record on this term's decided cases — it ingests the others, so it can't be benchmarked against them. Crowd = FantasySCOTUS; markets = Polymarket / Kalshi.
When the lower court builds on eroded precedent
A pattern we found while testing whether cited case law predicts reversal: when the court below leaned on a precedent the Supreme Court had already overruled, the Court reversed at a strikingly higher rate. It's a legible tell — the ground was visibly cracked before the decision — even though it's too rare to sharpen the model itself.
| Case (term) | Leaned on a precedent that was already overruled | Outcome |
|---|---|---|
| RIVERS v. GUERRERO (2024) | Teague v. Lane — overruled 2021 by Edwards v. Vannoy | Affirmed |
| CAMERON v. EMW WOMEN’S SURGICAL CENTER, P.S.C. (2021) | Apodaca v. Oregon — overruled 2020 by Ramos v. Louisiana | Reversed |
| DEVILLIER v. TEXAS (2023) | Williamson County Regional Planning Comm'n v. Hamilton Bank of Johnson City — overruled 2019 by Knick v. Township of Scott, Pennsylvania | Reversed |
| EDWARDS v. VANNOY, WARDEN (2020) | Harris v. United States — overruled 2013 by Alleyne v. United States | Affirmed |
| TOWN OF CHESTER v. LAROE ESTATES (2016) | McConnell v. Federal Election Commission — overruled 2010 by Citizens United v. FEC | Reversed |
| THOMPSON v. HEBDON (2019) | McConnell v. Federal Election Commission — overruled 2010 by Citizens United v. FEC | Reversed |
| MCCULLEN v. COAKLEY (2013) | Austin v. Michigan Chamber of Commerce — overruled 2010 by Citizens United v. FEC | Reversed |
| DARIN RYBURN, et al., PETITIONERS v. GEORGE R. HUFF, et al. (2011) | Saucier v. Katz — overruled 2009 by Pearson v. Callahan | Reversed |
| CHADRIN LEE MULLENIX, PETITIONER v. BEATRICE LUNA, INDIVIDUALLY AND AS REPRESENTATIVE OF THE ESTATE OF ISRAEL LEIJA, JR., et al. (2015) | Saucier v. Katz — overruled 2009 by Pearson v. Callahan | Reversed |
| WOOD v. MOSS (2013) | Saucier v. Katz — overruled 2009 by Pearson v. Callahan | Reversed |
| LOS ANGELES COUNTY, CALIFORNIA, PETITIONER v. CRAIG ARTHUR HUMPHRIES et al. (2010) | Saucier v. Katz — overruled 2009 by Pearson v. Callahan | Reversed |
| CHARLES A. REHBERG, PETITIONER v. JAMES P. PAULK (2011) | Saucier v. Katz — overruled 2009 by Pearson v. Callahan | Affirmed |
| HERNANDEZ v. MESA (2016) | Saucier v. Katz — overruled 2009 by Pearson v. Callahan | Reversed |
| JOHN D. ASHCROFT, PETITIONER v. ABDULLAH AL-KIDD (2010) | Saucier v. Katz — overruled 2009 by Pearson v. Callahan | Reversed |
| STEVE A. FILARSKY, PETITIONER v. NICHOLAS B. DELIA. (2011) | Saucier v. Katz — overruled 2009 by Pearson v. Callahan | Reversed |
"Recently-eroded" = overruled within 20 years before the case — and recency is the whole tell: leaning on a precedent overruled longago tracks the baseline, not this. Whether a precedent had been overruled is public before the Court rules, so it's a fair pre-decision signal. VotePredictor SCOTUS still doesn't use it as an input, for two reasons we measured: the model already reads these as reversal-prone from the case's posture (it calls them right ~88% of the time without this feature), and they're rare — 72 of 1169cases — too few to move the overall forecast either way. Shown here as an honest pattern, not a model feature. "Reversed" ≈ the Court sided with the petitioner (the side that appealed).
How this works
A gradient-boosted model predicts, for every justice on a case, the probability they side with the petitioner, using only pre-decision features (issue area, jurisdiction, the lower court's ruling and its ideological direction, cert reason, case origin) plus that justice's own prior record and the Court's composition. A separate case-level model produces the calibrated win probability. Both are trained walk-forward — to forecast a term, only earlier terms are used — so the accuracy above is out-of-sample. Vote data: Supreme Court Database(Spaeth et al.), the standard source for this work. "Ideology" in the profiles is the share of conservative votes, a descriptive proxy rather than a Martin–Quinn ideal point.