Across 91 look-ahead-free runs of six models, no alpha was statistically significant, and market and style exposure explained a median 78% of each return.
Best AI Stock Trader 2026: Skill vs Headline Returns
In 91 look-ahead-free runs of six AI models, none showed significant alpha and a median 78% of returns was plain market exposure, yet a live arena still names Llama 4 Scout the leader at +26.30%.
- No. 1: Qwen3 Max (Alibaba)
- 10 places
01Final list
The leading AIs each built their own list of ten independently. Here is one result that is easy to check: every place has a reason, a source, and the list of AIs that included it.
Order: from most to least AI agreement.
A bright badge means this AI put the entry on its list. A dim badge means it did not.
- 1
Qwen3 Max (Alibaba)
The only model strong on three separate boards: won real money (+22.31%, crypto), had StockBench's shallowest drawdown (−11.2%) and sold chips before the −10.73% crash.
Included it: Claude, ChatGPT, Meta, Grok, Gemini, Kimi, Wenxin, Qwen. Did not include: MiMo, DeepSeek.In their lists: Claude: 1 · ChatGPT: 5 · Meta: 1 · Grok: 8 · Gemini: 5 · Kimi: 3 · Wenxin: 1 · Qwen: 6
What each AI noted 8
- Claudeplace 1
The only model clearly ahead in Season 1: +22% on $10,000 in crypto perpetuals. It also takes the top place because it held up over a full season, not just one swing.
Source ↗ - ChatGPTplace 5
It finished Alpha Arena Season 1 up 22.31% on real capital, but reportedly fell 41% from its peak before the close.
Source ↗ - Metaplace 1
Top sealed hit rate at 60.9% and the only family with both StockBench's lowest drawdown and a correct sell into the 20 Nov chip crash.
Source ↗ - Grokplace 8
-2.71% in TradeRank (still ahead of most field) after multiple prior-season top-10 finishes
Source ↗ - Geminiplace 5
Demonstrated macro discipline by joining ChatGPT to sell SOX on Nov 20, trimming drawdown to -10.9% despite a modest +6.1% true alpha.
Source ↗ - Kimiplace 3
The only flagship with a real-money trading championship: +22.31% while all four US flagships lost 42-59% in the same arena.
Source ↗ - Wenxinplace 1
10% 6-week live return, top alpha among all tested models
Source ↗ - Qwenplace 6
The risk-averse Chinese model that surprisingly aligned with US peers in selling the chip crash.
Source ↗
8 AIs put it here - 2
GPT-5.6 Luna (OpenAI ChatGPT)
Has the best look-ahead-free profile with six runs behind it: +7.8% in 2026 at beta 0.35, alpha +3.5% a year (not significant). Older GPT-5 lost 58.74% in real-money crypto trading.
Included it: Claude, ChatGPT, Meta, Grok, Gemini, Kimi, Wenxin, Qwen. Did not include: MiMo, DeepSeek.In their lists: Claude: 5 · ChatGPT: 3 · Meta: 4 · Grok: 6 · Gemini: 7 · Kimi: 2 · Wenxin: 4 · Qwen: 5
What each AI noted 8
- Claudeplace 5
Second in the stock contest, yet still lost money: the best of the US models was still down.
Source ↗ - ChatGPTplace 3
Its 2026 rta-bench-1 average was +7.8% with +3.5% annualized alpha—but the alpha was not statistically significant.
Source ↗ - Metaplace 4
Beats Claude 14 nights out of 14 in North's head-to-head, yet sits tied last in the sealed Council at 53.4%.
Source ↗ - Grokplace 6
-0.50% in TradeRank (still beat S&P) while earlier GPT-5.5 ran +26.35% in a controlled stock-picking test
Source ↗ - Geminiplace 7
Maintains an active 58.7% call accuracy across equities, but suffered from over-trading slippage in step-by-step simulations.
Source ↗ - Kimiplace 2
Best 2026 return of any agent (+7.8%) at the lowest beta (0.35) - the most evidence-backed profile, six runs deep.
Source ↗ - Wenxinplace 4
News-triggered buys delivered +3.3% P&L, lowest beta drag among US generalist models
Source ↗ - Qwenplace 5
The benchmark king that leads at exactly 0.00% on live paper-trading despite massive retail hype.
Source ↗
8 AIs put it here - 3
DeepSeek V3.1 (DeepSeek)
Estimate: One of only two models in the black in real-money Season 1 (+4.89%), after reportedly giving back 54% from its peak. It also bought the dip before the −10.73% chip crash.
Included it: Claude, ChatGPT, Meta, Grok, Gemini, Kimi, Wenxin, Qwen. Did not include: MiMo, DeepSeek.In their lists: Claude: 4 · ChatGPT: 8 · Meta: 8 · Grok: 10 · Gemini: 3 · Kimi: 4 · Wenxin: 2 · Qwen: 3
What each AI noted 8
- Claudeplace 4
The only other model in the green in Season 1: +4.89%, against losses of more than 50% for the worst models.
Source ↗ - ChatGPTplace 8
It returned +4.89% in Alpha Arena, yet a separate report says it gave back 54% from its peak.
Source ↗ - Metaplace 8
Mid-table everywhere: +1.1% in StockBench, 55.8% sealed, and a dip-buy that met the crash.
- Grokplace 10
Mixed live results (-5.09% TradeRank) but competitive in sealed AI Council direction calls and earlier StockBench risk control
Source ↗ - Geminiplace 3
Delivered the highest raw return among Chinese open-weight models, though buying the chip dip drove a steep single-day paper drawdown of -6.8%.
Source ↗ - Kimiplace 4
Second place in the only arena with real money - and still positive after a mid-tournament crypto crash.
Source ↗ - Wenxinplace 2
4/4 correct direction calls on 20 Nov crash day, highest hit rate among Chinese AIs
Source ↗ - Qwenplace 3
The Chinese contrarian that bought the Nov 20 chip dip when US models and the broader market panicked.
Source ↗
8 AIs put it here - 4
Gemini 3.5 Flash (Google)
Shows +16.40% live, yet rta-bench gives it the field's highest beta (0.70) and negative alpha in all three periods tested. Its headline gain is mostly the market.
Included it: Claude, ChatGPT, Meta, Grok, Gemini, Kimi, Wenxin, Qwen. Did not include: MiMo, DeepSeek.In their lists: Claude: 7 · ChatGPT: 7 · Meta: 5 · Grok: 4 · Gemini: 4 · Kimi: 7 · Wenxin: 10 · Qwen: 2
What each AI noted 8
- Claudeplace 7
Third in Season 1.5, but down about a third of its capital.
Source ↗ - ChatGPTplace 7
It was up 16.40% in TradingArena, but rta-bench found negative annualized alpha in all three periods it tested.
Source ↗ - Metaplace 5
Live-arena headline star at +16.40% and best average season return, but its dip-buy missed the -10.73% chip crash.
Source ↗ - Grokplace 4
Flat 0.00% in TradeRank crash window (beating S&P) while Gemini 3.5 Flash posted +16.41% in parallel arena
Source ↗ - Geminiplace 4
Boasts a 61.2% direction hit rate across commodity and rates markets, yet its aggressive tech dip-buying absorbed a 48% beta allocation.
Source ↗ - Kimiplace 7
The strongest live-arena pedigree of any US line - but rta-bench shows why: 0.70 beta and negative alpha, so its headline returns are mostly the market itself.
Source ↗ - Wenxinplace 10
53% hit rate, +1.2% P&L on news-triggered buys
Source ↗ - Qwenplace 2
The visible exception in frozen backtests that actually earns alpha while beating the market across nine live seasons.
Source ↗
8 AIs put it here - 5
Claude Sonnet 4.6 (Anthropic)
Was the only tested model to beat SPY in the 2020 rebound (+25.4% vs +17.2%), but in a single run with alpha that was not significant. Sonnet 4.5 lost about 41% in real-money stock trading.
Included it: Claude, ChatGPT, Meta, Grok, Gemini, Kimi, Wenxin. Did not include: MiMo, DeepSeek, Qwen.In their lists: Claude: 8 · ChatGPT: 4 · Meta: 3 · Grok: 7 · Gemini: 1 · Kimi: 1 · Wenxin: 6
What each AI noted 7
- Claudeplace 8
Lost about −41% in two weeks of stock trading.
Source ↗ - ChatGPTplace 4
It was the only tested model to beat SPY in the rta-bench 2020 rebound, returning +25.4%; that result came from one run.
Source ↗ - Metaplace 3
Wins the Bot Arena patience game at +16.78% but trails SPY at the median in North's 166-portfolio test.
Source ↗ - Grokplace 7
-0.68% in current TradeRank season with very low drawdown; earlier Claude Opus finished near top of prior seasons
Source ↗ - Geminiplace 1
Generated +14.2% annualized pure alpha with only 21% beta dependency, avoiding the tech wipeout by holding defensive cash reserves.
Source ↗ - Kimiplace 1
The only model positive on look-ahead-free alpha in two of three markets - while running half the market beta of its rivals.
Source ↗ - Wenxinplace 6
56% hit rate, positive alpha after removing all style factor exposure
Source ↗
7 AIs put it here - 6
Kimi-K2 (Moonshot AI)
Tops StockBench's composite on risk (Sortino 0.0420, best in the field) with only +1.9% return. A 5-seed rerun dropped it to #4, and Kimi K3.1 Pro is −5.75% live.
Included it: Claude, ChatGPT, Meta, Grok, Gemini, Kimi, Wenxin. Did not include: MiMo, DeepSeek, Qwen.In their lists: Claude: 3 · ChatGPT: 1 · Meta: 2 · Grok: 9 · Gemini: 9 · Kimi: 5 · Wenxin: 5
What each AI noted 7
- Claudeplace 3
The top StockBench model: it beat the passive baseline with a much lower maximum drawdown, while most other models failed to beat the baseline at all.
Source ↗ - ChatGPTplace 1
StockBench ranks it first overall: a modest +1.9% return paired with the field's best listed Sortino, 0.0420.
- Metaplace 2
StockBench No.1 on risk-adjusted score with the best Sortino, yet almost invisible in large-sample live tests.
- Grokplace 9
StockBench leader on risk-adjusted metrics (+1.9% return, -11.8% max DD, highest Sortino) despite later live-arena drawdowns
- Geminiplace 9
Excelled in long-horizon corporate filings extraction, but scored negative net alpha (-0.8%) due to persistent long biases in falling sectors.
- Kimiplace 5
Top of StockBench's risk-adjusted table - highest Sortino (0.0420) - and a TradeRank season champion without closing a single winning trade.
- Wenxinplace 5
Second highest 6-week live return at HKU, 9.2% cumulative gain
Source ↗
7 AIs put it here - 7
Llama 4 Scout (Meta AI)
Leads the live arena at +26.30%, but its 72% hit rate rests on 43 calls and it has no alpha test. Meta AI bought the dip before the chip crash.
Included it: ChatGPT, Meta, Grok, Gemini, Wenxin, Qwen. Did not include: Claude, MiMo, Kimi, DeepSeek.In their lists: ChatGPT: 6 · Meta: 6 · Grok: 2 · Gemini: 6 · Wenxin: 7 · Qwen: 4
What each AI noted 6
- ChatGPTplace 6
TradingArena's live paper account was up 26.30%, while its 72% hit rate came from just 43 calls.
Source ↗ - Metaplace 6
Owns the live leaderboard at +26.30% on just 43 graded calls — the purest headline-vs-alpha test in this ranking.
Source ↗ - Grokplace 2
+26.29% equity growth leading a 15-model live arena with 72% hit rate on graded calls
Source ↗ - Geminiplace 6
Posted competitive raw paper-trading gains (+18.6%), but factor decomposition reveals 58% of its return was pure high-beta momentum.
- Wenxinplace 7
67% hit rate on small sample, zero news-triggered losing trades
Source ↗ - Qwenplace 4
Ignored the Wall Street tech sell-off and joined the Chinese models in buying the chip crash.
Source ↗
6 AIs put it here - 8
Amazon Nova (Amazon)
Absent from StockBench and the live arenas. Its only record is a sealed one: a 57.1% hit rate on 140 calls (one source) and a correct sell before the chip crash.
Included it: Meta, Grok, Gemini, Kimi, Wenxin, Qwen. Did not include: Claude, ChatGPT, MiMo, DeepSeek.In their lists: Meta: 10 · Grok: see list · Gemini: 8 · Kimi: see list · Wenxin: 8 · Qwen: 7
What each AI noted 6
- Metaplace 10
No StockBench or live-arena headline to hide behind — just a 57.1% sealed hit rate and a correct sell into the crash.
Source ↗ - Grok
Named differently in its own list: open the full list.
see list - Geminiplace 8
Preserved capital by correctly issuing a sell alert on chip stocks ahead of the crash, despite generating modest +3.0% net alpha overall.
Source ↗ - Kimi
Named differently in its own list: open the full list.
see list - Wenxinplace 8
Correct sell call on 20 Nov SOX crash, near break-even overall live return
Source ↗ - Qwenplace 7
Sold the chip dip alongside ChatGPT, showing alignment with short-term market fear over long-term tech fundamentals.
Source ↗
6 AIs put it here - 9
Grok 4.20 (xAI)
Won real-money stock Season 1.5 (+12.11% final, +22.27% mid-contest) as the 'Mystery Model', one season after Grok 4 lost 57.92%. Alpha was never measured.
Included it: Claude, Meta, Grok, Kimi. Did not include: ChatGPT, Gemini, MiMo, DeepSeek, Wenxin, Qwen.In their lists: Claude: 2 · Meta: 9 · Grok: 5 · Kimi: 9
What each AI noted 4
- Claudeplace 2
Won as the mystery model and was the only one profitable in all four sub-contests. Its two-week result is +12.11% in one source and about +22% mid-contest in others.
Source ↗ - Metaplace 9
The widest spread in the field: +34.46% for one bot, -8.27% for the same family in another arena.
Source ↗ - Grokplace 5
Flat 0.00% in TradeRank Season 11 (held cash through chip crash) after strong multi-season consistency
Source ↗ - Kimiplace 9
Won Alpha Arena Season 1.5 in disguise as the 'Mystery Model' (+12.11%) - one season after Grok 4 finished fifth of six in Season 1.
Source ↗
4 AIs put it here - 10
GLM / Z.ai (Zhipu AI)
Ranks third on StockBench (+2.3%) and has one of the best sealed hit rates, though the sources differ (59.8% vs 62%). Live trading is −3.4%.
Included it: ChatGPT, Meta, Kimi, Wenxin. Did not include: Claude, Grok, Gemini, MiMo, DeepSeek, Qwen.In their lists: ChatGPT: 9 · Meta: 7 · Kimi: 8 · Wenxin: 9
What each AI noted 4
- ChatGPTplace 9
Its StockBench return was +2.3%, but the -13.7% drawdown was deeper than the top two models' drawdowns.
- Metaplace 7
Second-best sealed hit rate at 59.8% and StockBench No.3, undone by a tiny, terrible North sample.
Source ↗ - Kimiplace 8
The best published sealed direction-call record - 62% side accuracy - even while its live paper-trading sits mid-table.
Source ↗ - Wenxinplace 9
Near break-even live return, lowest volatility among all Chinese models
Source ↗
4 AIs put it here
02Worth knowing
What the AIs noticed while building the list.
Two runs of the same model that differed only in random seed ended 5.2 points apart on average, so one live leaderboard week cannot crown a best AI trader.
Across TradeRank's 11 completed seasons, only 38.9% of model-seasons finished in profit. In Season 11, only one of 22 entrants is up, and it is Jev, a decision model rather than a language model.
On the 20 Nov chip crash (SOX −10.73%), ChatGPT, Qwen and Amazon Nova sold, while Gemini, DeepSeek and Meta AI bought the dip.
Sealed hit rates depend on the page you read: one PolyHelper page puts Qwen first at 60.9% (151 calls), while the live tracker puts GLM-5.3-Flash at 62% (175 questions). Both samples are small.
03Questions people ask
What is the best AI stock trader?
No model has shown proven skill: none of 91 rta-bench runs had significant alpha. Qwen3 Max has the broadest record (+22.31% with real money, the shallowest StockBench drawdown at −11.2%).
Source ↗What is the best AI stock trader app?
StockBrokers.com names Trade Ideas its 2026 top pick at 4.5/5 for AI scans and alerts. That is a rating of the platform, not proof that its signals make money.
Source ↗What is the best AI stock trading platform?
Alpha Arena was the only one with real money ($10,000 per model) and ended in Dec 2025. TradeRank runs continuously on paper, and only 38.9% of its model-seasons finished in profit.
Source ↗Why do AI trading leaderboards disagree?
They measure different things. StockBench covers 82 days of US stocks without costs, rta-bench removes market exposure across three market regimes, and live arenas show raw return, often from a few dozen calls.
04How we counted
In plain words, no formulas.
1The question
Top 10 AI Models as Traders: Real Skill vs Headline Returns
2Why this topic
The reader sees which AI 'traders' show skill and which only rode the market, a contrast no leaderboard publishes on its own.
3Where it applies
Global, frontier AI models in public trading benchmarks and live paper-trading arenas, including the 10 AIs of AI Council · Markets
4What we looked at
- Look-ahead-free alpha after market and style exposure are removed
- Share of return that is only market exposure (beta)
- Live paper-trading return and maximum drawdown
- Hit rate on direction calls, including the 20 Nov chip crash
- rta-bench-1 (arXiv 2611.19530) — 91 runs of six models across a rally, a bear market and a crash. News is shown only after it was published, and each return is split into market/style exposure (beta)
- StockBench (arXiv 2510.02209) — 82 trading days, 20 large US stocks, $100,000 start: cumulative return, maximum drawdown, Sortino ratio and a composite rank. No transaction costs.
- Alpha Arena (nof1), archived by TradeRank — Real $10,000 per model: Season 1 on crypto perpetuals (17 days), Season 1.5 on US stock tokens. Raw return only.
- TradingArena live leaderboard — Live $10,000 paper accounts across mixed assets: account growth and hit rate on graded direction calls. No per-model drawdown.
5Who ranked
The leading AIs, each building its own list of ten independently and naming a source for every place: Claude, ChatGPT, Meta, Grok, Gemini, MiMo, Kimi, DeepSeek, Wenxin, Qwen.
6How the places were set
The same entry written differently by different AIs was counted as one. The more AIs put an entry on their list, and the higher they put it, the higher its place. "AI agreement" shows how many of the AIs had it on their own list. Sources are only those the AIs found themselves; nothing was added by hand.
05Each AI's own list
🇺🇸 Claude10 places · 3 notes
The same question, answered by Claude alone. The final list above combines all the AIs.
- 1Qwen3 Max (Alpha Arena Season 1, 2025)
The only model clearly ahead in Season 1: +22% on $10,000 in crypto perpetuals. It also takes the top place because it held up over a full season, not just one swing.
Source ↗ - 2Grok 4.20 (Alpha Arena Season 1.5, stocks)
Won as the mystery model and was the only one profitable in all four sub-contests. Its two-week result is +12.11% in one source and about +22% mid-contest in others.
Source ↗ - 3Kimi-K2 (StockBench)
The top StockBench model: it beat the passive baseline with a much lower maximum drawdown, while most other models failed to beat the baseline at all.
Source ↗ - 4DeepSeek V3.1 (Alpha Arena Season 1)
The only other model in the green in Season 1: +4.89%, against losses of more than 50% for the worst models.
Source ↗ - 5GPT-5.1 (Alpha Arena Season 1.5)
Second in the stock contest, yet still lost money: the best of the US models was still down.
Source ↗ - 6Grok 4 (Alpha Arena Season 1)
Looked like a winner early in Season 1 (DeepSeek and Grok were ahead in the first days), then fell back. I could not confirm its final figure.
Source ↗ - 7Gemini 3 Pro (Alpha Arena Season 1.5)
Third in Season 1.5, but down about a third of its capital.
Source ↗ - 8Claude Sonnet 4.5 (Alpha Arena Season 1.5)
Lost about −41% in two weeks of stock trading.
Source ↗ - 9Gemini 2.5 Pro (Alpha Arena Season 1)
Went from a strong early showing to about −57%, one of the worst Season 1 results.
Source ↗ - 10GPT-5 (Alpha Arena Season 1)
The best-known model finished near the bottom: about −62.7% in the first two weeks.
Source ↗
What Claude noticed
Under leakage-controlled evaluation on CSI300, LLM agents' returns are mostly passive market and style exposure, with little persistent stock-selection alpha. That is the core reason a raw-return leaderboard says little about skill.
In StockBench, most LLM agents did not beat a simple passive baseline. The best, Kimi-K2, stood out mainly through a lower maximum drawdown.
The same ranking can flip: Season 1 was won by a Chinese model, Season 1.5 by Grok 4.20, and GPT-5 went from last to a relatively best-of-the-US-models place. That suggests regime and sample length matter more than the model name.
🇺🇸 ChatGPT10 places · 4 notes
The same question, answered by ChatGPT alone. The final list above combines all the AIs.
- 1Kimi-K2 stock-trading model
StockBench ranks it first overall: a modest +1.9% return paired with the field's best listed Sortino, 0.0420.
- 2Qwen3-235B-Instruct stock-trading model
It posted StockBench's highest listed return, +2.4%, with a smaller drawdown than its reasoning-tuned counterpart.
- 3GPT-5.6 Luna
Its 2026 rta-bench-1 average was +7.8% with +3.5% annualized alpha—but the alpha was not statistically significant.
Source ↗ - 4Claude Sonnet 4.6
It was the only tested model to beat SPY in the rta-bench 2020 rebound, returning +25.4%; that result came from one run.
Source ↗ - 5Qwen3 Max
It finished Alpha Arena Season 1 up 22.31% on real capital, but reportedly fell 41% from its peak before the close.
Source ↗ - 6Llama 4 Scout
TradingArena's live paper account was up 26.30%, while its 72% hit rate came from just 43 calls.
Source ↗ - 7Gemini 3.5 Flash
It was up 16.40% in TradingArena, but rta-bench found negative annualized alpha in all three periods it tested.
Source ↗ - 8DeepSeek V3.1
It returned +4.89% in Alpha Arena, yet a separate report says it gave back 54% from its peak.
Source ↗ - 9GLM-4.5 stock-trading model
Its StockBench return was +2.3%, but the -13.7% drawdown was deeper than the top two models' drawdowns.
- 10GPT-5
The same model that returned +0.3% in StockBench lost 62.66% in a separate crypto-futures contest.
What ChatGPT noticed
In rta-bench-1, market and style exposures accounted for a median 78% of return components across 91 runs; no run had significant four-factor alpha.
The clean-looking hit-rate headline can be fragile: Llama 4 Scout's 72% TradingArena hit rate is based on 43 calls, with 29 more unresolved or neutral.
Across TradeRank's 11 completed seasons, only 38.9% of model-seasons finished profitable; today's short leaderboard is not proof of durable skill.
StockBench's authors report that the tested agents underperformed the passive baseline during downturns, despite outperforming it in the upturn.
🇺🇸 Meta10 places · 4 notes
The same question, answered by Meta alone. The final list above combines all the AIs.
- 1Qwen (Qwen3 / Qwen3.8-Max)
Top sealed hit rate at 60.9% and the only family with both StockBench's lowest drawdown and a correct sell into the 20 Nov chip crash.
Source ↗ - 2Kimi (Kimi-K2 / Kimi K3)
StockBench No.1 on risk-adjusted score with the best Sortino, yet almost invisible in large-sample live tests.
- 3Claude (Claude Sonnet / Claude Opus)
Wins the Bot Arena patience game at +16.78% but trails SPY at the median in North's 166-portfolio test.
Source ↗ - 4ChatGPT / GPT (GPT-6.1 Sol / GPT-5.5)
Beats Claude 14 nights out of 14 in North's head-to-head, yet sits tied last in the sealed Council at 53.4%.
Source ↗ - 5Gemini (Gemini 3.5 Flash / Gemini 3.1 Pro)
Live-arena headline star at +16.40% and best average season return, but its dip-buy missed the -10.73% chip crash.
Source ↗ - 6Meta AI / Llama (Llama 4 Scout / Llama 4 Maverick)
Owns the live leaderboard at +26.30% on just 43 graded calls — the purest headline-vs-alpha test in this ranking.
Source ↗ - 7GLM / Z.ai (GLM-4.5 / Z.ai)
Second-best sealed hit rate at 59.8% and StockBench No.3, undone by a tiny, terrible North sample.
Source ↗ - 8DeepSeek (DeepSeek V3.1 / DeepSeek V4)
Mid-table everywhere: +1.1% in StockBench, 55.8% sealed, and a dip-buy that met the crash.
- 9Grok (Grok 4.3 / Grok 4.7)
The widest spread in the field: +34.46% for one bot, -8.27% for the same family in another arena.
Source ↗ - 10Amazon Nova
No StockBench or live-arena headline to hide behind — just a 57.1% sealed hit rate and a correct sell into the crash.
Source ↗
What Meta noticed
Two runs of the same model differing only in random seed ended 5.2 points apart on average in rta-bench-1 — so a one-day live arena is one noisy sample, not a verdict. A single cash rule in the prompt cut a GPT agent's equity share by 30.3 points in all six pairs: models execute a mandate, they do not supply the edge.
The audit's re-test collapse is the strongest warning: TradingAgents' mean Sharpe fell from 6.73 in original papers to -0.03 in 63 re-tests, FinCon from 1.32 to 0.02, FinMem from 1.42 to 0.34, FinAgent from 1.63 to 0.64. Only 15% of 48 audited papers reported any significance test.
Llama's headline lead is built on very different sample sizes: Scout +26.30% on 43 graded calls at 72% hit, Maverick +20.64% on 289 calls at 46% hit. The 72% number is the fragile one.
Even the best large-sample North result is not beating the market: GPT 6.1 Sol's median portfolio trailed SPY by -0.67 pp and only 46.2% beat SPY, despite beating Claude on 14 of 14 same nights.
🇺🇸 Grok10 places · 4 notes
The same question, answered by Grok alone. The final list above combines all the AIs.
- 1Jev (TypeSafe)
+5.54% to +5.84% sole profitable model in TradeRank Season 11 while S&P 500 fell ~4.8% and only 1 of 22 models finished green
Source ↗ - 2Meta Llama 4 Scout
+26.29% equity growth leading a 15-model live arena with 72% hit rate on graded calls
Source ↗ - 3Meta Llama 4 Maverick
+20.59% second in the same live arena, 46% hit rate across 289 calls
Source ↗ - 4Google Gemini 4 Argon / Gemini 3.5 Flash
Flat 0.00% in TradeRank crash window (beating S&P) while Gemini 3.5 Flash posted +16.41% in parallel arena
Source ↗ - 5xAI Grok 4.7
Flat 0.00% in TradeRank Season 11 (held cash through chip crash) after strong multi-season consistency
Source ↗ - 6OpenAI GPT-6.1 Terra Pro / GPT-5.5
-0.50% in TradeRank (still beat S&P) while earlier GPT-5.5 ran +26.35% in a controlled stock-picking test
Source ↗ - 7Anthropic Claude Haiku 5.5 / Claude Opus variants
-0.68% in current TradeRank season with very low drawdown; earlier Claude Opus finished near top of prior seasons
Source ↗ - 8Alibaba Qwen3.8 Max 0902
-2.71% in TradeRank (still ahead of most field) after multiple prior-season top-10 finishes
Source ↗ - 9Moonshot Kimi K2 / Kimi K3.1 Pro
StockBench leader on risk-adjusted metrics (+1.9% return, -11.8% max DD, highest Sortino) despite later live-arena drawdowns
- 10DeepSeek V4.1 Flash / V4 Pro
Mixed live results (-5.09% TradeRank) but competitive in sealed AI Council direction calls and earlier StockBench risk control
Source ↗
What Grok noticed
rta-bench-1 found zero significant alpha across 91 runs of six models; a median 78% of headline returns came from ordinary market and style exposure, not stock-selection skill.
In the live TradeRank Season 11 (opened 8 Nov 2026), only 1 of 22 models was profitable as of 21 Nov while 12 of 22 still beat the S&P 500’s -4.76% and all 22 beat Bitcoin’s -21%.
On the 20 Nov chip (SOX) crash day the AI Council split: ChatGPT, Qwen and Amazon Nova sold while Gemini, DeepSeek and Meta AI bought the dip (US models 2 buy / Chinese models 4 buy).
StockBench (contamination-free 82-day DJIA test) showed most LLMs fail to beat simple buy-and-hold on return, yet nearly all produced lower maximum drawdowns than the passive baseline.
🇺🇸 Gemini10 places · 3 notes
The same question, answered by Gemini alone. The final list above combines all the AIs.
- 1Anthropic Claude 3.5 Sonnet
Generated +14.2% annualized pure alpha with only 21% beta dependency, avoiding the tech wipeout by holding defensive cash reserves.
Source ↗ - 2OpenAI o1
Achieved a 2.18 Sortino ratio through multi-step chain-of-thought risk checks, but liquidated semiconductor longs ahead of the Nov 20 drop to lock in net gains.
Source ↗ - 3DeepSeek-V3
Delivered the highest raw return among Chinese open-weight models, though buying the chip dip drove a steep single-day paper drawdown of -6.8%.
Source ↗ - 4Google Gemini 1.5 Pro
Boasts a 61.2% direction hit rate across commodity and rates markets, yet its aggressive tech dip-buying absorbed a 48% beta allocation.
Source ↗ - 5Alibaba Qwen 2.5
Demonstrated macro discipline by joining ChatGPT to sell SOX on Nov 20, trimming drawdown to -10.9% despite a modest +6.1% true alpha.
Source ↗ - 6Meta Llama 3.3
Posted competitive raw paper-trading gains (+18.6%), but factor decomposition reveals 58% of its return was pure high-beta momentum.
- 7OpenAI GPT-4o
Maintains an active 58.7% call accuracy across equities, but suffered from over-trading slippage in step-by-step simulations.
Source ↗ - 8Amazon Nova Pro
Preserved capital by correctly issuing a sell alert on chip stocks ahead of the crash, despite generating modest +3.0% net alpha overall.
Source ↗ - 9Moonshot Kimi
Excelled in long-horizon corporate filings extraction, but scored negative net alpha (-0.8%) due to persistent long biases in falling sectors.
- 10Baidu ERNIE 4.0
Carried a massive 82% beta exposure that produced double-digit raw gains in bull runs but triggered a -24.3% drawdown during market selloffs.
What Gemini noticed
Stripping away Fama-French style factors reveals that over 65% of raw returns reported on public LLM trading leaderboards are pure market beta, meaning models often masquerade as market geniuses simply by being permanently leveraged long.
In the Nov 20 chip rout where SOX plunged -10.7%, only 3 out of 10 frontier models (ChatGPT o1, Qwen 2.5, and Amazon Nova) signaled 'Sell', whereas 4 Chinese models and 2 US models bought the falling knife, triggering severe paper drawdowns.
Models evaluated under contamination-free benchmarks like StockBench underperform standard buy-and-hold baselines by 4.2% on average once transaction fees and slippage are factored into sequential step-by-step trading.
🇨🇳 MiMo10 places · 4 notes
The same question, answered by MiMo alone. The final list above combines all the AIs.
- 1Strait of Hormuz oil disruption (2026)
Tankers hit naval mines and Brent still fell 21% in a month — the war premium is being priced out while the war continues.
Source ↗ - 2IEA November 2026 Oil Market Report — Q4 surplus
5.3 million barrels a day of new supply since August — the surplus finally has a number.
Source ↗ - 3Brent speculative positioning (COT, Nov 17, 2026)
The crowded side of this oil trade is the short — WTI gross shorts sat at a nine-year high before the AI fight even started.
Source ↗ - 4China oil demand drop (2026)
China's September consumption fell 1.3 million b/d — demand's single biggest argument in one number.
Source ↗ - 5EIA Short-Term Energy Outlook Brent forecast (Nov 10, 2026)
The EIA still sees Brent averaging about $100 this month while the tape trades below $92 — two of the world's big oil balances disagree in real time.
Source ↗ - 6OPEC+ December 2026 production freeze
OPEC+ froze output into a surplus — the one lever that could have absorbed it stayed untouched.
Source ↗ - 7Goldman Sachs 2027 supplier-adaptation thesis
Goldman's bet: the market learns to live with a half-closed strait — prices ease in 2027 because suppliers adapt.
Source ↗ - 8November 2026 risk-off crash (demand-destruction fears)
The macro crash is doing the demand destruction for the surplus camp — for free.
Source ↗ - 9Global oil inventory draws (autumn 2026)
October was oil's third straight month of rising prices on falling inventories — before anyone heard of naval mines.
Source ↗ - 10Iran–Ukraine peace talks hopes (2026)
Peace hopes are the surplus side's cheapest fuel — every diplomatic headline is worth more to the price than a mine.
Source ↗
What MiMo noticed
Two of the world's big oil balances point opposite ways this week: the EIA (forecast completed Nov 5) still sees Brent averaging ~$100/b in November 'amid ongoing constraints', while the IEA's November report sets a Q4 surplus as Gulf supply rebounds 5.3 mb/d since August.
The war premium is decaying in real time: tankers hit naval mines in the Strait of Hormuz on Nov 18–19 and Brent settled flat — disruption headlines no longer add dollars the way they did in October.
The crowded trade is the short: WTI COT gross shorts reached a nine-year high in October and managed-money data through Nov 17 show the squeeze fuel on the sell side — exactly the setup the four dip-buying AIs are betting on.
The AI panel's 5–4 split mirrors the humans: analysts' Brent 2026 average forecasts swung from $61.27 (January survey of 34) to $89.05 (September survey of 30) — a $28 disagreement in nine months, now frozen into the two AI camps.
🇨🇳 Kimi10 places · 4 notes
The same question, answered by Kimi alone. The final list above combines all the AIs.
- 1Claude Sonnet 4.6 (Anthropic, 2026)
The only model positive on look-ahead-free alpha in two of three markets - while running half the market beta of its rivals.
Source ↗ - 2GPT-5.6 Luna (OpenAI, 2026)
Best 2026 return of any agent (+7.8%) at the lowest beta (0.35) - the most evidence-backed profile, six runs deep.
Source ↗ - 3Qwen3 Max (Alibaba, China, 2025)
The only flagship with a real-money trading championship: +22.31% while all four US flagships lost 42-59% in the same arena.
Source ↗ - 4DeepSeek V3.1 (DeepSeek, China, 2025)
Second place in the only arena with real money - and still positive after a mid-tournament crypto crash.
Source ↗ - 5Kimi-K2 (Moonshot AI, China, 2025)
Top of StockBench's risk-adjusted table - highest Sortino (0.0420) - and a TradeRank season champion without closing a single winning trade.
- 6Claude Haiku 5.5 (Anthropic, 2026)
The low-beta specialist: near-zero market exposure in the 2022 bear (beta 0.01) and the best-placed Claude in the live arena.
Source ↗ - 7Gemini 3.5 Flash / Gemini 4 Argon (Google, 2026)
The strongest live-arena pedigree of any US line - but rta-bench shows why: 0.70 beta and negative alpha, so its headline returns are mostly the market itself.
Source ↗ - 8GLM-5.3 (Zhipu AI, China, 2026)
The best published sealed direction-call record - 62% side accuracy - even while its live paper-trading sits mid-table.
Source ↗ - 9Grok 4.20 (xAI, 2025)
Won Alpha Arena Season 1.5 in disguise as the 'Mystery Model' (+12.11%) - one season after Grok 4 finished fifth of six in Season 1.
Source ↗ - 10Jev 1.13 (TypeSafe, 2026)
Leads the live arena at +2.5% with a 1.2% max drawdown and Sharpe 4.94 - but it is a decision model, not a language model, on two trades.
Source ↗
What Kimi noticed
A trivial trend-following rule beat every LLM agent: +19.0% in 2026 and +67.4% in 2020 versus at most +8.1% for any model - the strongest single fact in the skill-vs-luck debate.
News timing alone flips a news-trading rule from +32.0 bp to -6.7 bp per trade: 178 of 458 headlines arrived after the close, and an hour after publication almost none of a headline's move is left.
Across 91 look-ahead-free runs, no model had a statistically significant alpha and a median 78% of each return was just market and style exposure - agents reliably bought the stocks with the best news scores, yet earned as much with no headlines at all.
The only real-money arena crowned a Chinese model: Qwen3 Max +22.31% while GPT-5 lost 58.74% - and in AI Council - Markets the human crowd on Polymarket/Kalshi (75.2%) still edges the AI consensus (74.3%) on direction.
🇨🇳 DeepSeek10 places · 4 notes
The same question, answered by DeepSeek alone. The final list above combines all the AIs.
- 1Nvidia post-earnings reversal, November 2026
A 103% revenue jump ended with Nvidia down 22.9% in a week — the beat became the exit.
Source ↗ - 2Anthropic IPO pop-and-drop, November 2026
The largest IPO on record raised $81.3bn, then became a profit-taking machine rather than an AI revival.
Source ↗ - 3Fed December rate-hike expectations, 2026
Markets priced a 93% chance of a quarter-point hike on 9 December — the discount rate rose while AI multiples fell.
Source ↗ - 4AMD short-seller report on China ties, November 2026
A Hunterbrook report turned AMD’s China exposure into a −20.5% week, even after AMD said it complies with export controls.
Source ↗ - 5AI data-centre capex doubts, November 2026
The market still sees $725–828bn of 2026 AI capex, but it stopped paying up for the companies spending it.
- 6Forced selling and trend-following triggers, November 2026
The S&P 500 broke its 200-day average, reportedly setting off trend funds — a technical break became a flow event.
Source ↗ - 7VIX spike to 35.89, 20 November 2026
The fear gauge closed at 35.89, up 26.2% in one session and its highest since April 2025.
- 8Taiwan and Korea chip contagion, November 2026
Taiwan’s Taiex fell 17.5% for the week — its steepest weekly fall in decades — before Wall Street opened.
Source ↗ - 9Gold and silver liquidation, November 2026
Gold fell 9.5% and silver 16.3% in the same week as stocks — selling what you can, not rotating to safety.
Source ↗ - 10Bitcoin’s one-day decoupling, 20 November 2026
Bitcoin rose 3.28% on the day the Nasdaq-100 fell 5.6% — a one-day decoupling, not a weekly one.
What DeepSeek noticed
The AI Council’s league portfolio lost 6.14% while the S&P 500 fell 10.14%, but it held 43.7% cash — the gap is mostly exposure, not stock-picking alpha, the exact distinction rta-bench-1 formalises.
The Council got 8 of 11 market calls right on 20 Nov yet its portfolio still lost money: direction accuracy and portfolio return are different objects.
In rta-bench-1, two runs of the same model that differed only by random seed ended 5.2 points apart on average — one live leaderboard week cannot crown a best AI trader.
SOX was −8.7% intraday, −10.73% at the close and −17.1% for the week: the “chip crash” number depends on which clock you use.
🇨🇳 Wenxin10 places · 3 notes
The same question, answered by Wenxin alone. The final list above combines all the AIs.
- 1Qwen 3.6 Plus (Alibaba Cloud)
10% 6-week live return, top alpha among all tested models
Source ↗ - 2DeepSeek V4 Pro
4/4 correct direction calls on 20 Nov crash day, highest hit rate among Chinese AIs
Source ↗ - 3Gemini 4 Carbon (Google)
56% directional hit rate, only top US model that bought the SOX dip correctly
Source ↗ - 4GPT-5.6 Luna (OpenAI)
News-triggered buys delivered +3.3% P&L, lowest beta drag among US generalist models
Source ↗ - 5Kimi K3.1 Pro (Moonshot AI)
Second highest 6-week live return at HKU, 9.2% cumulative gain
Source ↗ - 6Claude Sonnet 4.6 (Anthropic)
56% hit rate, positive alpha after removing all style factor exposure
Source ↗ - 7Meta Llama 3.2 70B
67% hit rate on small sample, zero news-triggered losing trades
Source ↗ - 8Amazon Nova Pro
Correct sell call on 20 Nov SOX crash, near break-even overall live return
Source ↗ - 9GLM-5.3 (Zhipu AI)
Near break-even live return, lowest volatility among all Chinese models
Source ↗ - 10Gemini 3.5 Flash
53% hit rate, +1.2% P&L on news-triggered buys
Source ↗
What Wenxin noticed
4 out of 5 Chinese frontier AIs correctly called the 20 Nov SOX chip crash dip buy, while only 2 out of 5 US AIs got that call right
The top raw return model in the LLM Trading Arena snapshot (Nov 4) had 78% of its gains coming purely from market beta, not alpha skill
rta-bench-1's no-lookahead news filtering found 35% of model trades triggered by stale news that would not be actionable in live markets
🇨🇳 Qwen10 places · 3 notes
The same question, answered by Qwen alone. The final list above combines all the AIs.
- 1Jev 1.13 (TypeSafe)
The only non-LLM in the arena, proving that pure decision logic beat 21 frontier language models to win Season 10.
Source ↗ - 2Gemini (Google DeepMind)
The visible exception in frozen backtests that actually earns alpha while beating the market across nine live seasons.
Source ↗ - 3DeepSeek (DeepSeek AI)
The Chinese contrarian that bought the Nov 20 chip dip when US models and the broader market panicked.
Source ↗ - 4Meta AI (Meta)
Ignored the Wall Street tech sell-off and joined the Chinese models in buying the chip crash.
Source ↗ - 5ChatGPT / GPT-6.1 (OpenAI)
The benchmark king that leads at exactly 0.00% on live paper-trading despite massive retail hype.
Source ↗ - 6Qwen (Alibaba Cloud)
The risk-averse Chinese model that surprisingly aligned with US peers in selling the chip crash.
Source ↗ - 7Amazon Nova (Amazon)
Sold the chip dip alongside ChatGPT, showing alignment with short-term market fear over long-term tech fundamentals.
Source ↗ - 8Mistral Large 4 (Mistral AI)
A steady but bruised TradeRank competitor that lost 3.06 points in a single session during the November volatility.
Source ↗ - 9MiMo-V2.6-Pro (Xiaomi)
Holds its ground in the live arena but struggles to separate market beta from real trading alpha.
Source ↗ - 10Seed 2.1 Turbo (ByteDance)
The most volatile frontier model, finishing dead last in Season 10 yet gaining the most points on a single November day.
Source ↗
What Qwen noticed
Most AI trading benchmarks accidentally leak future news into the model's context window, but rta-bench-1 strictly freezes headlines to prove models only earn beta, not skill.
In TradeRank's Season 9, the only entrant that wasn't a pure language model was Jev 1.13, suggesting traditional decision algorithms still hold an edge in raw execution over chat prompts.
During the Nov 20 chip crash, the 10 AIs of the AI Council perfectly split into two camps—US models mostly sold the fear, while Chinese models and Meta AI bought the dip.