Lab: the whole full page →

Before you trust an AI with money, see how it actually advised

Four virtual accounts right now:

AIs beating the index: 0 of 14; AIs beating the crowd: 0 of 14.

Council-minus-crowd interval: no conclusion yet, 86 questions of the 350 needed.

Every piece of advice is sealed before the market opens, then checked against the outcome.

Dispute of the day

Fear index (VIX), day: 71% of US AIs say up, 43% of Chinese AIs

Fear index (VIX)Gold10-year US Treasury yieldCopper

Since the last issue: Nasdaq-100 — the turned to a fall; Chip index — the consensus turned to a fall; 10-year US Treasury yield — the consensus turned to a rise; Fear index (VIX) — the consensus turned to a rise. Changed side on the day: ChatGPT (chip index, fear index (VIX)); Qwen (chip index); Kimi (fear index (VIX)); Grok (chip index); DeepSeek (Nasdaq-100, chip index, fear index (VIX)); Meta AI (S&P 500, Nasdaq-100, chip index).

Past forecasts scored since the last issue: Week from Wed, Nov 4, 2026 — the consensus was right on 5 of 10; Wed, Nov 11, 2026 — the consensus was right on 7 of 10.

What the Council would do

Forecasts of the leading AIs of the USA and China for 12 markets over four horizons. Where a model has no new answer, its position is unchanged.

IndicatorNowDayWeekMonth6 monthsAIs vs the crowd
S&P 500 · other and the crowd ▾7,635.30▲ +0.21%Hold · 71% of the AIs chose itaverage chance of a rise: USA 44% · China 44%(effective voices ≈ 2.0)
WeekHold · 79% of the AIs chose itaverage chance of a rise: USA 49% · China 44%(effective voices ≈ 1.8)
MonthTrim · 50% of the AIs chose itaverage chance of a rise: USA 51% · China 43%(effective voices ≈ 2.7)
6 monthsBuy on dips · 50% of the AIs chose itaverage chance of a rise: USA 60% · China 59%(effective voices ≈ 1.8 · on deviations from climatology ≈ 1.10)
AIs vs the crowdAIs: 40% chance of a rise · crowd 28% · effective voices ≈ 2.0 ↗ price at Thu 6:40 AM , $7k traded
Nasdaq-100 · other horizons and the crowd ▾28,433.53▲ +0.21%Trim · 57% of the AIs chose itaverage chance of a rise: USA 41% · China 41%(effective voices ≈ 2.0)
WeekHold · 57% of the AIs chose itaverage chance of a rise: USA 44% · China 42%(effective voices ≈ 1.8)
MonthTrim · 36% of the AIs chose itaverage chance of a rise: USA 52% · China 40%(effective voices ≈ 2.7)
6 monthsBuy on dips · 43% of the AIs chose itaverage chance of a rise: USA 61% · China 58%(effective voices ≈ 1.8 · on deviations from climatology ≈ 1.10)
AIs vs the crowdAIs: 38% chance of a rise · crowd 26% · effective voices ≈ 2.0 ↗ price at Thu 6:40 AM ET, $9 traded thin market
Dow Jones · other horizons and the crowd ▾52,682.52▲ +0.37%Hold · 86% of the AIs chose itaverage chance of a rise: USA 52% · China 52%(effective voices ≈ 2.0)
WeekHold · 71% of the AIs chose itaverage chance of a rise: USA 53% · China 53%(effective voices ≈ 1.8)
MonthHold · 57% of the AIs chose itaverage chance of a rise: USA 55% · China 55%(effective voices ≈ 2.7)
6 monthsBuy on dips · 57% of the AIs chose itaverage chance of a rise: USA 58% · China 60%(effective voices ≈ 1.8 · on deviations from climatology ≈ 1.10)
AIs vs the crowdno price read
Chip index · other horizons and the crowd ▾10,557.32▼ −0.064%Trim · 79% of the AIs chose itaverage chance of a rise: USA 41% · China 41%(effective voices ≈ 2.0)
WeekHold · 64% of the AIs chose itaverage chance of a rise: USA 45% · China 44%(effective voices ≈ 1.8)
MonthHold · 43% of the AIs chose itaverage chance of a rise: USA 55% · China 47%(effective voices ≈ 2.7)
6 monthsBuy on dips · 43% of the AIs chose itaverage chance of a rise: USA 64% · China 60%(effective voices ≈ 1.8 · on deviations from climatology ≈ 1.10)
AIs vs the crowdno market
10-year US Treasury yield · other horizons and the crowd ▾5.105▲ +10.7 bp · yield ↑ · bonds get cheaperToward a rise, on a dip · 50% of the AIs chose itaverage chance of a rise: USA 56% · China 53%(effective voices ≈ 2.0)
WeekToward a rise, on a dip · 50% of the AIs chose itaverage chance of a rise: USA 59% · China 56%(effective voices ≈ 1.8)
MonthKeep the bet · 43% of the AIs chose itaverage chance of a rise: USA 53% · China 55%(effective voices ≈ 2.7)
6 monthsToward a fall · 57% of the AIs chose itaverage chance of a rise: USA 40% · China 47%(effective voices ≈ 1.8 · on deviations from climatology ≈ 1.10)
AIs vs the crowdno market
Dollar index · other horizons and the crowd ▾103.18▲ +0.26%Hold · 93% of the AIs chose itaverage chance of a rise: USA 52% · China 53%(effective voices ≈ 2.0)
WeekHold · 43% of the AIs chose itaverage chance of a rise: USA 56% · China 56%(effective voices ≈ 1.8)
MonthHold · 50% of the AIs chose itaverage chance of a rise: USA 57% · China 54%(effective voices ≈ 2.7)
6 monthsTrim · 50% of the AIs chose itaverage chance of a rise: USA 44% · China 46%(effective voices ≈ 1.8 · on deviations from climatology ≈ 1.10)
AIs vs the crowdno price read
Fear index (VIX) · other horizons and the crowd ▾17.59▲ +0.57%Toward a rise, on a dip · 50% of the AIs chose itaverage chance of a rise: USA 56% · China 51%(effective voices ≈ 2.0)
WeekToward a fall · 29% of the AIs chose itaverage chance of a rise: USA 48% · China 49%(effective voices ≈ 1.8)
MonthKeep the bet · 36% of the AIs chose itaverage chance of a rise: USA 49% · China 52%(effective voices ≈ 2.7)
6 monthsKeep the bet · 57% of the AIs chose itaverage chance of a rise: USA 51% · China 50%(effective voices ≈ 1.8 · on deviations from climatology ≈ 1.10)
AIs vs the crowdno market
Gold · other horizons and the crowd ▾3,876.20▼ −3.11%Hold · 57% of the AIs chose itaverage chance of a rise: USA 44% · China 42%(effective voices ≈ 2.0)
WeekHold · 50% of the AIs chose itaverage chance of a rise: USA 49% · China 42%(effective voices ≈ 1.8)
MonthHold · 36% of the AIs chose itaverage chance of a rise: USA 55% · China 44%(effective voices ≈ 2.7)
6 monthsBuy on dips · 43% of the AIs chose itaverage chance of a rise: USA 62% · China 59%(effective voices ≈ 1.8 · on deviations from climatology ≈ 1.10)
AIs vs the crowdAIs: 37% chance of a rise · crowd 11% · effective voices ≈ 2.0 ↗ price at Thu 6:40 AM ET, $575 traded thin market
WTI crude oil · other horizons and the crowd ▾87.63▲ +1.27%Hold · 57% of the AIs chose itaverage chance of a rise: USA 41% · China 43%(effective voices ≈ 2.0)
WeekHold · 43% of the AIs chose itaverage chance of a rise: USA 41% · China 41%(effective voices ≈ 1.8)
MonthTrim · 50% of the AIs chose itaverage chance of a rise: USA 44% · China 40%(effective voices ≈ 2.7)
6 monthsHold · 64% of the AIs chose itaverage chance of a rise: USA 49% · China 50%(effective voices ≈ 1.8 · on deviations from climatology ≈ 1.10)
AIs vs the crowdAIs: 41% chance of a rise · crowd 37% · effective voices ≈ 2.0 ↗ price at Thu 6:40 AM ET, $1k traded another window: this question ends Thu 6:00 PM ET, not at our event's end
Brent crude oil · other horizons and the crowd ▾94.53▲ +1.91%Hold · 57% of the AIs chose itaverage chance of a rise: USA 43% · China 43%(effective voices ≈ 2.0)
WeekHold · 43% of the AIs chose itaverage chance of a rise: USA 40% · China 41%(effective voices ≈ 1.8)
MonthTrim · 43% of the AIs chose itaverage chance of a rise: USA 44% · China 40%(effective voices ≈ 2.7)
6 monthsHold · 64% of the AIs chose itaverage chance of a rise: USA 48% · China 49%(effective voices ≈ 1.8 · on deviations from climatology ≈ 1.10)
AIs vs the crowdno market
Copper · other horizons and the crowd ▾6.194▼ −1.52%Hold · 71% of the AIs chose itaverage chance of a rise: USA 48% · China 46%(effective voices ≈ 2.0)
WeekHold · 57% of the AIs chose itaverage chance of a rise: USA 52% · China 45%(effective voices ≈ 1.8)
MonthHold · 36% of the AIs chose itaverage chance of a rise: USA 58% · China 49%(effective voices ≈ 2.7)
6 monthsBuy more · 43% of the AIs chose itaverage chance of a rise: USA 64% · China 59%(effective voices ≈ 1.8 · on deviations from climatology ≈ 1.10)
AIs vs the crowdno market
Bitcoin · other horizons and the crowd ▾71,730▼ −3.65%Hold · 57% of the AIs chose itaverage chance of a rise: USA 44% · China 45%(effective voices ≈ 2.0)
WeekHold · 50% of the AIs chose itaverage chance of a rise: USA 44% · China 41%(effective voices ≈ 1.8)
MonthTrim · 36% of the AIs chose itaverage chance of a rise: USA 52% · China 42%(effective voices ≈ 2.7)
6 monthsBuy on dips · 43% of the AIs chose itaverage chance of a rise: USA 60% · China 56%(effective voices ≈ 1.8 · on deviations from climatology ≈ 1.10)
AIs vs the crowdAIs: 44% chance of a rise · crowd 50% · effective voices ≈ 2.0 ↗ price at Thu 6:40 AM ET, $20 traded another window: this Polymarket question ends Fri 12:00 PM ET, not at our event's end thin market

Lab: this block exists only in the full version →

The AIs against the crowd

A close question people bet real money on at and : the AIs’ own odds of “Up” and the crowd’s price (each market’s day is in the card). A daily crowd question exists for 5 of the 12 markets. A crowd question counts from its own reference — on Kalshi, above the previous close at 4:00 PM; the score of the day counts from the open.

The same question as the crowd (the very market, its rules and its end):

For orientation — not quite the same question (another window, or Kalshi's price read between the two nearest strikes):

In one line: the AIs argue moderately, are neutral overall, their confidence is low.

34
0 — all agree, 100 — an even split
citations of a site another AI also cites: 46%, of the same link: 6% · cited most: reuters.com 21%
?Models can rely on the same news. Agreement is not independent confirmation; the heat shows how far their actions differ.
Direction
47%
mildly bearish · the AIs' average odds of a rise today over the 12 markets · effective voices ≈ 2.0
Confidence
13
How far the AIs' day odds are from 50/50, on average: 0 — a coin, 100 — all certain.
Each model on its own, by : Day 13 · Week 15 · Month 16 · 6 months 18.
They said 57% — it came true 55% (2186 day forecasts over 18 sessions, since Oct 19, 2026).

Right now · what this is and the scores

Where the virtual accounts stand, highest first: holding cash $100,258 · a 60/40 stock/bond portfolio $99,594 · the S&P 500 (bought and held) $98,885 · the Council $97,069.

Measured on 91 comparable questions over 3 weeks; a conclusion needs 350. This is the standing, not a verdict.

The crowd is more accurate than the Council by 2.2 points (crowd 75.8%, Council 73.6%), 91 questions.

Council minus crowd, share of right sides: 95% interval from −11.0 to 3.9 points, 86 questions.

Still noise, no conclusion drawn: 91 questions so far.

The Council's virtual portfolio from the day you pick

Virtual $100,000 put into the Council's portfolio at the close before the chosen day, against the same sum in the S&P 500 bought and held; the league's own costs are inside.

  • Oct 19: By Nov 11: the Council $97,069 · the S&P 500 $98,885 (from the Oct 19 open, first price 09:30–09:45 ET, after costs)
  • Oct 20: By Nov 11: the Council $97,246 · the S&P 500 $99,130 (from the Oct 19 close, the Council and the S&P 500 alike, without the entry cost)
  • Oct 21: By Nov 11: the Council $97,037 · the S&P 500 $98,799 (from the Oct 20 close, the Council and the S&P 500 alike, without the entry cost)
  • Oct 22: By Nov 11: the Council $96,715 · the S&P 500 $98,282 (from the Oct 21 close, the Council and the S&P 500 alike, without the entry cost)
  • Oct 23: By Nov 11: the Council $97,756 · the S&P 500 $99,457 (from the Oct 22 close, the Council and the S&P 500 alike, without the entry cost)
  • Oct 26: By Nov 11: the Council $97,977 · the S&P 500 $99,979 (from the Oct 23 close, the Council and the S&P 500 alike, without the entry cost)
  • Oct 27: By Nov 11: the Council $97,842 · the S&P 500 $99,274 (from the Oct 26 close, the Council and the S&P 500 alike, without the entry cost)
  • Oct 28: By Nov 11: the Council $97,531 · the S&P 500 $98,720 (from the Oct 27 close, the Council and the S&P 500 alike, without the entry cost)
  • Oct 29: By Nov 11: the Council $97,334 · the S&P 500 $98,381 (from the Oct 28 close, the Council and the S&P 500 alike, without the entry cost)
  • Oct 30: By Nov 11: the Council $97,610 · the S&P 500 $98,759 (from the Oct 29 close, the Council and the S&P 500 alike, without the entry cost)
  • Nov 2: By Nov 11: the Council $98,422 · the S&P 500 $100,541 (from the Oct 30 close, the Council and the S&P 500 alike, without the entry cost)
  • Nov 3: By Nov 11: the Council $98,775 · the S&P 500 $101,025 (from the Nov 2 close, the Council and the S&P 500 alike, without the entry cost)
  • Nov 4: By Nov 11: the Council $98,987 · the S&P 500 $100,724 (from the Nov 3 close, the Council and the S&P 500 alike, without the entry cost)
  • Nov 5: By Nov 11: the Council $98,958 · the S&P 500 $99,087 (from the Nov 4 close, the Council and the S&P 500 alike, without the entry cost)
  • Nov 6: By Nov 11: the Council $99,948 · the S&P 500 $103,658 (from the Nov 5 close, the Council and the S&P 500 alike, without the entry cost)
  • Nov 9: By Nov 11: the Council $99,956 · the S&P 500 $103,335 (from the Nov 6 close, the Council and the S&P 500 alike, without the entry cost)
  • Nov 10: By Nov 11: the Council $99,501 · the S&P 500 $100,946 (from the Nov 9 close, the Council and the S&P 500 alike, without the entry cost)
  • Nov 11: By Nov 11: the Council $100,096 · the S&P 500 $101,319 (from the Nov 10 close, the Council and the S&P 500 alike, without the entry cost)
Paper alpha fund (virtual): rules published, start 2026-11-09, $100,000

Rules, published before the count: a virtual fund of $100,000. Every trading day it goes long or short on the Council's consensus (the mean forecast weighted by distance from 50%) across the markets, equal weights inside the long and the short half, rebalanced on Mondays, result by each market's daily move. Compared with the S&P 500 and with cash.

Collected 0 of 30 resolved positions; no conclusion before that.

This week's book (set on Monday)

  • Brent crude oil: short
  • Bitcoin: short
  • Copper: short
  • Dow Jones: long
  • Dollar index: long
  • Gold: short
  • Nasdaq-100: short
  • Chip index: short
  • S&P 500: short
  • 10-year US Treasury yield: long
  • Fear index (VIX): long
  • WTI crude oil: short

Observation, not a conclusion. Virtual money only.

The league of virtual accounts: the AIs and the crowd side by side

The league. Session Wed, Nov 11, 2026: traded — DeepSeek, Kimi · answered and kept the weights — ChatGPT, MiMo (Xiaomi), Gemini, Z.ai, Grok, Claude, Qwen, Meta AI, StepFun · positions unchanged — Nemotron, Amazon Nova, ERNIE (Wenxin).

Marks: ▲ up · ▼ down · ● even (the heat map); ● traded at the session · ○ kept its weights (the league).

Every AI keeps three separate virtual $100,000 accounts, so one AI has three figures: the League (buying only), the Pro league ( and leverage up to 2×) and the portfolio by actions (from its calls on the markets).

Both leagues in short: each AI’s money and status

League‡

Money: each AI manages a virtual $100,000, buying only — who has made more.

Starting line, not a ranking: no AI has 30 closed forecasts yet.

ParticipantEquityChangeStatus
S&P 500, bought once and held since Mon, Oct 19 · from the Oct 19 open, first price 09:30–09:45 , after costs$98,885−$1,115−1.12%—
1 ChatGPTUSA · Council$98,740−$1,260−1.26%○ kept the weights · new decision accepted, takes effect Thu, Nov 12
2 MiMo (Xiaomi)China · Council$98,317−$1,683−1.68%○ kept the weights · new decision accepted, takes effect Fri, Nov 13
3 StepFunChina · bench$98,018−$1,982−1.98%○ kept the weights · new decision accepted, takes effect Fri, Nov 13
4 GeminiUSA · Council$97,982−$2,018−2.02%○ kept the weights · new decision accepted, takes effect Fri, Nov 13
5 Z.aiChina · Council$97,812−$2,188−2.19%○ kept the weights · new decision accepted, takes effect Fri, Nov 13
6 NemotronUSA · bench$97,751−$2,249−2.25%○ holds its positions
7 Amazon NovaUSA · bench$96,820−$3,180−3.18%○ holds its positions · new decision accepted, takes effect Thu, Nov 12
8 GrokUSA · Council$96,818−$3,182−3.18%○ kept the weights · new decision accepted, takes effect Thu, Nov 12
9 DeepSeekChina · Council$96,658−$3,342−3.34%● traded: 7 trades · new decision accepted, takes effect Thu, Nov 12
10 ClaudeUSA · Council$96,583−$3,417−3.42%○ kept the weights · new decision accepted, takes effect Thu, Nov 12
11 KimiChina · Council$96,581−$3,419−3.42%● traded: 7 trades · new decision accepted, takes effect Thu, Nov 12
12 QwenChina · Council$95,631−$4,369−4.37%○ kept the weights · new decision accepted, takes effect Thu, Nov 12
13 Meta AIUSA · Council$95,287−$4,713−4.71%○ kept the weights · new decision accepted, takes effect Thu, Nov 12
14 ERNIE (Wenxin)China · bench$94,225−$5,775−5.78%○ holds its positions · new decision accepted, takes effect Thu, Nov 12
More rows — the whole table ↓

Pro league‡

the same AIs and money, plus shorts and up to 2× leverage · leader Gemini +$2,922 (+2.92%) · S&P 500 since Tue, Oct 20 −0.85%
ParticipantEquityChangeStatus
1 Gemini$102,922+$2,922+2.92%○ kept the weights · new decision accepted, takes effect Fri, Nov 13
S&P 500, bought once and held since Tue, Oct 20 · from the Oct 20 open, first price 09:30–09:45 ET, after costs$99,154−$846−0.85%—
2 StepFun$98,404−$1,596−1.60%○ kept the weights · new decision accepted, takes effect Fri, Nov 13
3 MiMo (Xiaomi)$97,596−$2,404−2.40%○ kept the weights · new decision accepted, takes effect Fri, Nov 13
4 DeepSeek$97,418−$2,582−2.58%● traded: 9 trades · new decision accepted, takes effect Thu, Nov 12
5 Nemotron$97,238−$2,762−2.76%○ holds its positions
6 ChatGPT$97,225−$2,775−2.78%● traded: 6 trades · new decision accepted, takes effect Thu, Nov 12
7 Grok$96,626−$3,374−3.37%○ kept the weights · new decision accepted, takes effect Thu, Nov 12
8 Z.ai$96,377−$3,623−3.62%○ kept the weights · new decision accepted, takes effect Fri, Nov 13
9 Amazon Nova$95,686−$4,314−4.31%● traded: 6 trades · new decision accepted, takes effect Fri, Nov 13
10 Claude$95,168−$4,832−4.83%○ kept the weights · new decision accepted, takes effect Thu, Nov 12
11 Qwen$94,457−$5,543−5.54%● traded: 8 trades · new decision accepted, takes effect Thu, Nov 12
12 ERNIE (Wenxin)$94,351−$5,649−5.65%● traded: 9 trades · new decision accepted, takes effect Thu, Nov 12
13 Meta AI$92,760−$7,240−7.24%○ kept the weights · new decision accepted, takes effect Thu, Nov 12
14 Kimi$92,105−$7,895−7.90%○ kept the weights · new decision accepted, takes effect Thu, Nov 12

The league’s full table, with the Pro league →

League cards (PNG, 1200×630): the portfolio league · the Pro league

Lab: this block exists only in the full version →

Can you trust the AI chorus

Herd index

78%

How alike the AIs' probabilities are today (100% = they all say the same).

When all the AIs agree too much, that is a reason to doubt.

Where the AIs disagreed most

  1. Chip index: 33% (DeepSeek) vs 70% (nova)
  2. Fear index (VIX): 35% (Wenxin (文心, Baidu; ERNIE)) vs 75% (nova)
  3. 10-year US Treasury yield: 30% (nova) vs 61% (ChatGPT)

Prompt drift: how much the answer moves with the wording

BRENT: spread 4.65 points across 4 wordings (from 42% to 51%).

BTC: spread 4.79 points across 4 wordings (from 40% to 50%).

COPPER: spread 6.29 points across 4 wordings (from 38% to 51%).

A wide spread means the answer depends on the wording, so it should be trusted less.

Calibration: best ECE 0.022 (nova) · AIs overconfident: 2 of 14
  • nova: ECE 0.022 · Brier 0.239 · underconfident: sounds 59% sure, right 63%
  • DeepSeek: ECE 0.028 · Brier 0.235 · underconfident: sounds 56% sure, right 60%
  • stepfun: ECE 0.048 · Brier 0.239 · underconfident: sounds 54% sure, right 63%
  • Grok: ECE 0.050 · Brier 0.236 · underconfident: sounds 57% sure, right 62%
  • nemotron: ECE 0.050 · Brier 0.258 · overconfident: sounds 57% sure, right 51%
  • mimo: ECE 0.062 · Brier 0.241 · underconfident: sounds 55% sure, right 59%
  • zai: ECE 0.062 · Brier 0.241 · underconfident: sounds 55% sure, right 61%
  • Claude: ECE 0.064 · Brier 0.244 · underconfident: sounds 53% sure, right 59%
  • Wenxin (文心, Baidu; ERNIE): ECE 0.072 · Brier 0.241 · honest confidence
  • Kimi: ECE 0.073 · Brier 0.248 · honest confidence
  • ChatGPT: ECE 0.073 · Brier 0.240 · honest confidence
  • Qwen: ECE 0.074 · Brier 0.240 · honest confidence
  • gemini: ECE 0.083 · Brier 0.245 · overconfident: sounds 59% sure, right 54%
  • Meta AI: ECE 0.148 · Brier 0.239 · honest confidence

Yardstick (ForecastBench): superforecasters 0.086, best AI models 0.103, a coin (always 50%) 0.25.

What is a good Brier score?
  • 0 is perfect: every call was certain and right.
  • 0.25 is the same as always saying 50%: no skill at all.
  • 1 is the worst: sure of the answer and wrong. Our table and the yardsticks above use the same scale, so they can be compared directly.
Who is right in which regime: Council Brier from 0.218 (oil (WTI) up 2% or more on the day) to 0.249 (VIX below 18), 6 regimes
  • VIX below 18: 44 questions · Council Brier 0.249 · best AI Claude, 0.244
  • VIX 18 to 25: 113 questions · Council Brier 0.233 · best AI Qwen, 0.225
  • oil (WTI) down 2% or more on the day: 32 questions · Council Brier 0.240 · best AI nova, 0.231
  • oil (WTI) up 2% or more on the day: 28 questions · Council Brier 0.218 · best AI Qwen, 0.195
  • S&P 500 up on the day: 101 questions · Council Brier 0.237 · best AI Grok, 0.228
  • S&P 500 down on the day: 58 questions · Council Brier 0.239 · best AI Qwen, 0.221

An observation, not a conclusion. Regime = the market's state in the question's own snapshot; Brier: 0 = perfect, 0.25 = a coin; the regimes overlap.

ECE = how far the stated odds are from reality (0 = perfect). Brier: 0 = perfect, 0.25 = a coin. Season 1 resolved forecasts, draws are not counted.

Book of misses: 25 of 142 cases where the Council and the crowd were both wrong
  1. 2026-11-05 · Dow Jones: Council 68%, crowd 87%, the market went down
    “The Dow Jones Industrial Average vaulted 3.03% yesterday to 53,356.12—adding over 1,575 points—and Dow futures are up…” — gemini
  2. 2026-10-20 · WTI crude oil: Council 62%, crowd 84%, the market went down
    “EIA weekly data released yesterday showed US crude inventories fell 7.2 million barrels, far larger than the 2 million…” — Wenxin (文心, Baidu; ERNIE)
  3. 2026-11-02 · : Council 68%, crowd 75%, the market went down

Court of the Council · one fact against the majority

After the seal, up to three AIs that disagreed with the majority name one fact it may be missing. The sealed forecast does not change.

  • ChatGPT on COPPER: Shanghai Futures Exchange warehouse stocks fell 24% in the week to November 6, complicating the bearish demand narrative. (2026-11-12). Would admit the majority was right at 6.20. source
  • Qwen on WTI: OPEC+ output surged 2.04 mb/d in October, completely offsetting geopolitical risk premiums. (2026-11-12). Would admit the majority was right at 85.98. source
  • Claude on BRENT: Iran's deputy armed-forces chief says Hormuz stays shut until the Supreme Leader orders it; speaker Ghalibaf says it will never return to pre-war status. (2026-11-12). Would admit the majority was right at 92.28. source

Lab: this block exists only in the full version →

How to verify

How to check it ▾

Check it in one step: one question, word for word, to every AI · sealed in our open hash-chained book, then anchored in Bitcoin · recount the score: CSV · the book · one news packet of the closed book

Sealed in our open hash-chained book before the outcome; the hash is anchored in Bitcoin (OpenTimestamps). Some answers were on a signed list before the open, the others came after it; those count from the first price after their answer — see proof of the day. From Oct 21: signed before the open; Oct 19–20: our own hash-chained book, anchored in Bitcoin. The book is the one open file of every sealed forecast, line by line.

Method and an honest frame

How this is built and what we compare it with

What studies showed and how we differ

  • A study of fine-tuned agents found their errors correlate at 0.70. Source
    How we differ: We do not borrow this figure: we measure the overlap between our own answers separately and show it on the page.
  • The PolyBench study, on live Polymarket data: 2 of 7 models ended in profit. Source
    How we differ: We use virtual money only and show every loss; our method differs, so the results are not directly comparable.
  • Polymarket lists about 600 markets on which model will lead LMArena. Source
    How we differ: For us this is only a possible source of questions, not a result and not proof of anyone's accuracy.
  • Prediction Arena, real-money trading: on Kalshi 6 models lost 22.6% on average, on Polymarket about 1.1%; Grok guessed 51.9% of markets. Source
    How we differ: We trade virtual money only, and we publish our losses too.
  • The Prophet Arena paper evaluated forecasters on 1,367 resolved real-world events. Source
    How we differ: We do not copy its setup; our questions, scoring and virtual money are our own, so the numbers are not comparable.
  • The Foresight Arena paper estimates that detecting a true edge of 0.02 at 80% power needs about 350 resolved binary predictions. Source
    How we differ: For us this is a reminder that small samples prove little; we show sample sizes next to results.
  • The FinBench pilot covered one trading day, three tickers and 33 forecasts. Source
    How we differ: A pilot this small shows little; we make no comparison with it.
  • A study of forecasting with retrieved online discussions tested on 216 resolved prediction markets. Source
    How we differ: We do not borrow its setup or numbers; we show our own results with sample sizes.
  • A paper on earnings-call mention markets evaluated 856 Kalshi markets. Source
    How we differ: Different markets and method; we only note that sample sizes there are large, and show ours next to results.

Who else does this, and how we differ

From our own research register. Outside projects are described only as the register describes them; where it says the substance is not checked, it is not checked.

Register item 71: Forecaster Arena, an open project, runs 7 models on Polymarket events; llm-predict-alpha works on short binary Polymarket markets. These are the closest comparisons; we study how they present results.

Register item 72: LightningRod offers a product for forecasting on prediction markets. Its substance is not checked.

Register item 73: Kalshi has an internal agent, Harrison, that reads news, watches competitors and proposes new contracts (Bloomberg, 15.06.2026). This matches our idea of a model proposing the day's questions.

Register item 77: Mantic raised $25 million (18.09.2026); a retelling says AI took 1st, 2nd and 5th place at the Metaculus Cup. Second-hand, dates differ between sources and are to be checked. It is an objection to our claims and we keep it in view.

Register item 78: Public launched agents for Kalshi (24.09.2026). Real money, not our concern; an agent acting on a probability trigger is an idea for a virtual league.

Register item 78: Polymarket closed its framework for agents (11.05.2026). Real money, not our concern.

Register item 45: Meta Arena, an internal Meta project (press reports, июнь 2026, second-hand): points only, no money, not released, no results. How we differ: ours is sealed, virtual $100,000, public.

Our difference: virtual money only, and the comparison is with the crowd's price on the same questions.

What speaks against us

Honest objections, each with what we check. Virtual money only.

  • In Alpha Arena Season 1.5 only 5 of 32 accounts ended in profit. We check: does our council stay ahead of the plain market baseline over many questions, not one good week.
  • In TradeRank season 2 only 3 of 13 agents made a profit, and all 3 were contrarian. We check: does a contrarian rule beat the crowd in our own sealed record.
  • Forecasting Research Institute: superforecasters score a Brier of 0.086, the best language models 0.103 (March 2026); parity is only an extrapolation. We compare with the crowd price, not with superforecasters, and we do not claim parity. We check the gap on our own sealed questions. ↗
  • The startup Mantic raised $25 million on 18 September 2026 after outscoring human entrants in a forecasting contest. Money raised is not proof of accuracy. We check every forecast with a seal made before the outcome.

Questions and answers

Who are these AIs?

Several independent AI models from different companies. Each gets the same question about every market, answers on its own, and every answer is shown under the model’s name.

Where does the data come from, and how can I check a forecast?

Prices are public market quotes. Every forecast is sealed with a timestamp and a hash before the New York open and counted against the real close afterwards. The sealed lists and all the data can be downloaded (CSV, JSON), so anyone can recount the score.

Are the AIs better than the market?

Not proven. So far 91 comparable questions are counted of the 350 needed for a statistical verdict; today the data do not show that the AIs beat the market.

Is this investment advice?

No. All money here is virtual. The forecasts are an experiment in how well AIs call markets, not a recommendation to buy or sell anything.

Daily digest subscription

Subscribe to the digest (on the main site) →