Hundreds of AI trading agents, six months of data, no edge

AI trading bots are usually judged by a leaderboard screenshot. A preprint posted to arXiv on September 4, 2026 takes a different route: it publishes the full behavior record of two fleets of language-model trading agents that ran in production, with real user configurations, over roughly six months.

Who ran it, and on what

The paper comes from DX Research Group (DXRG), the lab that built and operates both systems. That matters: the authors are reporting on their own products, and the paper has not been peer reviewed.

• DX Terminal Pro: 3,505 user-funded vaults, each run by a Qwen3-235B agent, traded real ETH in a 12-token memecoin market on the Base network for 21 days, from February 26 to March 18, 2026.
• DXAP live alpha: from June 8 to August 15, 2026, 500 to 599 user-created agents, with 91 to 117 active at any time, traded perpetual futures on Hyperliquid. Most accounts were paper accounts starting at $10,000 with live prices, plus a small real-money book.

Together, the record covers about 7.5 million model calls, roughly 300,000 onchain actions, 231,638 multi-tool turns and 14,596 fills.

The scorecard

The DXAP agents were compared with a matched sample of 1,961 Hyperliquid retail traders with $1,000 to $100,000 in equity. Over the window the authors report, the agents won 41% of round trips against 50% for retail nContext: n=353 agent and 16,009 retail round trips · 4 days to Aug 15, 2026 · DXRG preprint arXiv:2609.05663, Table 2 · margin: not disclosed. Only 15% of agents active during the week were net positive, against 53% of retail accounts nContext: n=117 agents and 4,828 retail accounts · week to Aug 15, 2026 · DXRG preprint arXiv:2609.05663, Table 2 · margin: not disclosed. Cumulative realized P&L for the fleet stood at minus $217,000, restated at a common fee of 5.5 basis points per side. The authors are explicit: "Neither fleet shows a directional edge."

On DX Terminal Pro, the median vault ended at 0.492 times its true starting value, and 16.2% of vaults finished in profit. nContext: n=3,505 vaults · Feb 26 to Mar 18, 2026 · DXRG preprint arXiv:2609.05663 · margin: not disclosed

Finding 1: the sliders drive the bot

Users configured agents with five sliders (trade frequency, risk tolerance, trade size, holding style, diversification) plus free-text strategy. The risk slider alone added about 0.425x of leverage per level.

What the agent was shown mattered as much as what it was told. DXAP agents saw a "movers" list of nine symbols. 46.5% of entries landed in those nine, against 8.9% if picks had been random nContext: n: entries not stated · DXAP live alpha, Jun to Aug 2026 · DXRG preprint arXiv:2609.05663 · interval 43.4 to 49.7 as published, type not stated. Symbols ranked just below the display cutoff were chosen far less often, even though their market conditions were statistically the same.

Finding 2: leverage ignores volatility

Across 6,400 closed positions, median leverage was 5.0x in every volatility bucket, even though volatility varied 5.7 times from the calmest bucket to the wildest. The liquidation rate climbed from 0.7% to 4.3%. nContext: n=6,400 closed positions · Jun 8 to Jul 24, 2026 · DXRG preprint arXiv:2609.05663, Table 3 · margin: not disclosed

Liquidations were concentrated: 128 of 205 (62%) came from one configuration, momentum-style agents at the maximum trade-frequency setting, which made up about 11% of the book. nContext: n=205 liquidations among 6,400 positions · Jun 8 to Jul 24, 2026 · DXRG preprint arXiv:2609.05663 · margin: not disclosed

A prompt fix did not work. When agents were required to state their liquidation distance before entering, the ones that stated it were liquidated more often, 5.8% versus 1.2%. nContext: n: not disclosed · DXAP live alpha, 2026 · DXRG preprint arXiv:2609.05663 · margin: not disclosed

Finding 3: they find the upside, then give it back

43.2% of positions were up at least 3% (300 basis points) at some point within 24 hours. Of those, 49.3% still closed at a loss. Attaching a fixed bracket at entry, a 2% stop and a 4% target, improved results by 39.0 basis points per position on average, mostly by cutting the worst losses. Even so, none of the 16 exit policies tested was profitable outright. nContext: n=6,400 closed positions, 2,765 of them up 3% or more · Jun 8 to Jul 24, 2026 · DXRG preprint arXiv:2609.05663 · bracket gain interval +21.3 to +56.5 bps, day-clustered

Finding 4: swapping the model did not help

The lab replayed 416 captured production scenarios through three frontier models. Their decision quality was statistically indistinguishable. Cost was not: running the full replay cost $8.25 with qwen3.7-plus against $203.61 with claude-fable-5, roughly a 25-fold difference.

Herding showed up too. On March 1, 1,544 of 3,454 active DX Terminal Pro vaults bought the same memecoin within one hour.

The rules the authors now follow

The paper ends with a 17-rule checklist for anyone evaluating trading agents, built, in the authors' words, on their own retractions: three of their earlier positive results were withdrawn. Examples include treating the market day rather than the single trade as the unit of evidence, and comparing results at a common fee rate.