Better than superforecasters? Why everyone is No. 1 in AI forecasting right now
Korean startup Neurofusion claims first place out of 550 models and says it beat top human forecasters. Cassi, DeepMind and others made similar claims before it. We explain how all of this can be true at once and what the numbers actually show.

What this is about
On November 20, 2026, several South Korean outlets, including Hellot and StartupN, reported on the financial AI startup Neurofusion. It reportedly ranks first out of 550 models on ForecastBench, a competition for AI forecasting, and also outperformed the human "superforecasters." Superforecasters are people who have been unusually accurate in forecasting tournaments over many years. But in July, the benchmark's operator, the Forecasting Research Institute (FRI) led by psychologist Philip Tetlock, had named Cassi AI the strongest system, alongside submissions from xAI and Google DeepMind. And it explicitly described the result only as a likely tie. Neurofusion had already announced a "first place" back in September, and Cassi calls itself the "#1 AI on the ForecastBench leaderboard" on its website.
How can so many players be in the lead at the same time? The answer lies in how the benchmark works. And it matters right now, because the FRI's own extrapolation predicted parity between AI and superforecasters for exactly this month: November 2026.
Three leaderboards, many cutoff dates
ForecastBench poses new yes/no questions about the future every two weeks. Scoring uses the Brier score, the squared difference between the stated probability and the actual outcome. Since March 2026, it has been converted into a "Brier Index": 100 means perfect foresight, 50 is what you'd get by answering "50 percent" on every question. There are two question types:
- Dataset questions are generated automatically from time series such as FRED, Yahoo Finance, ACLED or Wikipedia. Example: Will the policy rate go up?
- Market questions come from forecasting platforms such as Metaculus, Polymarket, Manifold and the RAND Forecasting Initiative. The FRI considers them the better test of judgment on novel, one-off events.
On top of that, there are three leaderboards. The preliminary leaderboard only scores dataset questions, starting from 225 resolved questions, after about ten days. The website warns: "Rankings are not finalized." The tournament leaderboard combines both question types equally, and a submission only appears there after 50 days. Finally, the baseline leaderboard shows only models that the FRI itself runs without any tools.
That's how the conflicting claims add up:
| Who claims "No. 1"? | Leaderboard | As of | AI score | Superforecasters |
|---|---|---|---|---|
| Cassi AI | Tournament, overall (best AI system, not best participant) | Jul 17, 2026 | 68.8 | 69.2 (1st place) |
| Neurofusion | Preliminary, dataset only | Sep 18, 2026 | 67.5 (of 464) | 63.6 |
| Neurofusion | Tournament, overall | Nov 17, 2026 | 69.4 (of 550) | 68.1 (9th place) |
Sources: Cassi website, ETNews, Hellot/StartupN. Neurofusion's figures are company claims. We couldn't access the live leaderboard and therefore couldn't verify them ourselves.
In September, ETNews also mentioned anonymous Neurofusion submissions in places 2 through 4. The "550 models" apparently count every submission separately, including test variants and the FRI's own runs. The Korean reports don't say where Cassi stands in November. They only say that among teams with disclosed identities, only Neurofusion and DeepMind (68.4) are ahead of the humans. The places in between are therefore anonymous.
The benchmark that moves even though nobody is forecasting anymore
The superforecasters last made forecasts in July 2024, on different questions than today's models. The FRI bridges that gap with a difficulty adjustment. For dataset questions, it uses a statistical model (two-way fixed effects) to estimate how hard each question is. For market questions, the market's own Brier score serves as the difficulty measure. If you get a hard question right, you earn more points.
That has an odd side effect, which we pieced together from several sources:
| Date | Superforecaster index (overall) |
|---|---|
| March 2026 | 70.6 |
| May 23, 2026 | 70.2 |
| July 6/17, 2026 | 69.3 / 69.2 |
| November 17, 2026 | 68.1 |
The humans haven't submitted a single new forecast in two years, yet their score keeps dropping. An author on the EA Forum did the math. The superforecasters' adjusted score has been unchanged at −0.0173 since July 2024. But their index on market questions fell by 4.7 points within four months, because the question pool got harder and shifted the reference point. In other words, the finish line moves too. Neurofusion's 1.3-point lead is in the same ballpark as these shifts.
We back-calculated the values. An index of 69.4 corresponds to a Brier score of about 0.094, an index of 68.1 to about 0.102. That's roughly 8 percent fewer errors. Neurofusion's claim of "15 to 16 percent fewer errors" refers to Claude Opus 5.5 (66.6) and GPT-6 Astra (66.7), not to the humans. That calculation checks out.
p-values around 0.4: what "tie" means
In its July post, the FRI tested whether the systems differ from the superforecasters. Cassi got a p-value of 0.41, xAI 0.16 and 0.15, DeepMind 0.14. The p-value tells you how easily a gap like this could arise by chance alone. By convention, you only speak of a real difference below 0.05. The confidence intervals, the range of plausible values, overlap heavily. That's why the FRI writes that the data is "more consistent with parity than superiority." Neurofusion publishes no p-value for its lead.
On dataset questions, the picture is clearer: in July, 17 submissions were ahead of the superforecasters there. According to Good Judgment, one superforecaster pointed to a fundamental problem. When dozens of systems compete every two weeks:
"With enough evaluations, one will eventually land below that mark purely by chance."
The Copier Line: copy the market price and you'll get far
For market questions, according to the ForecastBench paper, the models get the "freeze value" in their prompt: the crowd's forecast on the platform at the time the question set was created (ten days before the submission deadline). Back in 2025, the FRI found that GPT-4.5 essentially copied this value: the correlation was 0.994, and on 26 of 122 questions it reproduced the value exactly.
The EA Forum analysis "The Copier Line" from August 21, 2026, takes this to its logical conclusion. It calculates the score a participant would get by simply returning the price it was shown. On August 11, this Copier Line stood at 70.93 points. The results:
- Of 527 participants tested, none beat the market price significantly; 442 were significantly below it.
- Closest were Cassi (p = 0.004 without correction for multiple testing) and the superforecaster median (p = 0.017).
- 75 percent of the 273 leaderboard entries were at or below the Copier Line.
- The displayed price alone lifted the baseline models by a median of 15.7 points, from about 51.5 to 68.1.
The superforecasters were 4.67 points above the line, Cassi 5.67. So the AI's "win" on market questions means it's slightly better than the market price it was handed. Neurofusion's market score (73.9 vs. 73.5 for the humans) is 0.4 points above the humans'. The author also argues that the difficulty adjustment isn't neutral, because the market is easiest to beat on uncertain questions. His demand: publish the Copier Line next to every score.
The cross-checks: Metaculus and Good Judgment
Metaculus FutureEval measures directly, on the same questions, using yes/no, numeric and multiple-choice questions. In spring 2026, ten pros on 99 questions were on average 1.25 points ahead of a team of the ten best bots (confidence interval −4.87 to 2.37, p = 0.247). In 2025, the gap had still been 17.7 to 20 points. The summer results, merged on GitHub on October 15, showed just −0.26 points (p = 0.94). The pros won 60 of 97 questions, but on average it's a tie. On September 5, according to StackFutures, a bot won the Metaculus Cup for the first time.
Good Judgment sells superforecaster predictions and so has a commercial interest. Still, in "What ForecastBench Doesn't Measure" the company makes points you can verify. In 2024, the humans worked alone, under time pressure, with three to eight people per question instead of 16 or more. There are only yes/no questions. Teamwork, updating forecasts and writing good questions aren't tested. In its own client work, with a small sample, the company says it hasn't seen AI at superforecaster level. According to Good Judgment, a Financial Times test on Fed decisions showed the AI was roughly as good as the market but didn't beat it.
For comparison, the Bridgewater report on the AIA Forecaster: we checked the numbers in the paper. 0.1076 vs. 0.1110 for the superforecasters is correct, but measured retrospectively and on a single round of questions. On market questions, the humans were ahead (0.0740 vs. 0.0753), and on liquid markets the market consensus beat the AI (0.1106 vs. 0.1258).
Our verdict
Several companies can be "No. 1" because they're referring to different leaderboards, different cutoff dates and different comparison groups ("best AI," "best with a name"). Add to that a 2024 benchmark that shifts mathematically over time.
Our take, broken down by question type:
- Dataset questions (interest rates, prices, time series, short horizons): AI really is on par or ahead here, and not even Good Judgment disputes that.
- Market questions: Statistically, it's a tie. But the achievement mostly consists of slightly improving on a market price the model was given. The tournament doesn't show how good the systems would be without that price.
- Head-to-head on the same questions (Metaculus): The gap is practically closed, with no clear winner.
- Long horizons, sparse data, well-crafted questions, team forecasts: There's hardly any solid data, either for or against AI.
Notably, the FRI's extrapolation to November 2026 looks set to hold. The LEAP expert panel's median expectation was that AI wouldn't surpass superforecasters until 2030; the superforecasters themselves said 2028.
If you want to use AI forecasts for finance or politics, here's the takeaway. Where a liquid market exists, its price is the bar to beat, and no system has demonstrably beaten it yet. AI pays off mainly where there's no market and lots of data questions need fast answers. And a spot on a benchmark leaderboard says nothing about returns, as even StartupN concedes in its Neurofusion report. The company markets an investment platform called Valley AI, Cassi sells forecasts, and so does Good Judgment. The picture will only get clearer after the new superforecaster round the FRI has announced for fall 2026. That's when humans and machines will once again go head-to-head on the same questions.
Sources
- Hellot: Neurofusion, ForecastBench 550 Modelle Platz 1 hellot.net
- StartupN: Neurofusion Platz 1 unter 550 Modellen startupn.kr
- ETNews: Neurofusion Platz 1 auf vorläufiger Rangliste (Sept. 2026) m.etnews.com
- AI Market Watch: Neurofusion-Behauptung (Sept. 2026) ai-market-watch.com
- ForecastBench: Leaderboards forecastbench.org
- FRI Substack: AI models have likely reached parity with superforecasters forecastingresearch.substack.com
- FRI Substack: LLMs Are Closing the Gap on Human Superforecasters forecastingresearch.substack.com
- FRI Substack: How well can large language models predict the future? forecastingresearch.substack.com
- FRI Substack: Introducing the Brier Index forecastingresearch.substack.com
- arXiv: ForecastBench-Paper (v5) arxiv.org
- EA Forum (GreaterWrong-Spiegel): The Copier Line ea.greaterwrong.com
- Good Judgment: What ForecastBench Doesn't Measure (Yet) goodjudgment.substack.com
- Good Judgment: What Superforecasters Actually Said About ForecastBench goodjudgment.substack.com
- Good Judgment: Not So Fast goodjudgment.substack.com
- LessWrong: FutureEval Spring Results lesswrong.com
- GitHub: Metaculus/aib-analysis PR #33 (Summer 2026) github.com
- LessWrong: AI Forecasting in 2026 – What 11 Analyses Say lesswrong.com
- arXiv: AIA Forecaster Technical Report (Bridgewater) arxiv.org
- LEAP: ForecastBench performance leap.forecastingresearch.org
- Cassi AI cassi-ai.com
- Lightning Rod: AI Reaches Human SuperForecaster Range blog.lightningrod.ai
- StackFutures: AI wins Metaculus Cup stackfutures.com
- Telling the Future: Tetlock Among the Machines tellingthefuture.substack.com
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.