The bankroll

Does agent mode work?

There is no demonstrated edge. This page holds the evidence, including the two models that made money and the reasons the evidence against is weaker than its headlines suggest. None of it is our own measurement.

The short answer#

Three statements, and nothing stronger is supported by anything published:

There is no demonstrated edge yet.
In the one public competition with real capital, four of six frontier models finished down between 31% and 63% over about two weeks. The other two finished up, by 22.3% and 4.9%.
A 2026 preprint covering 7.5 million model invocations across two production fleets found no directional edge and a median 5.0x leverage in every volatility sextile, though it is the vendor reporting on its own products rather than an independent audit.

Every source here is secondary, and dated

Nothing below is a measurement this project made. Access date for every source on this page is 2026-09-22. Where a primary source could not be loaded, that is said explicitly rather than glossed.

Alpha Arena Season 1#

Run by nof1. A live competition, not a fund and not a study.

Fact	Value
Ran	2025-10-18 to 2025-11-04
Duration	17 days, so about two weeks
Stake	$10,000 of real capital per model
Venue	Hyperliquid perpetual futures
Setup	six frontier models, identical prompts and inputs, deciding every two to three minutes

Model	Final return
Qwen3 Max	+22.3%
DeepSeek V3.1	+4.89%
Claude Sonnet 4.5	-30.81%
Grok 4	-45.3%
Gemini 2.5 Pro	-56.71%
GPT-5	-62.66%

Four of six lost between 31% and 63%, and two of six finished up. Any summary that says only "poor" is leaving out the two winners, which a reader finds in one search.

The best performer traded least

Qwen3 Max made roughly 43 trades in total, fewer than three a day, across a competition where models were deciding every two to three minutes. Whatever that is evidence of, it is not evidence that more inference produces better returns.

What the Arena evidence is worth#

Source	What it gives
`iweaver.ai/blog/alpha-arena-ai-trading-season-1-results/`	the full final leaderboard and the trade-count detail
`protos.com/llm-crypto-trading-contest-finds-llms-cant-trade-crypto/`	independent secondary coverage; its headline framing is literally "LLM crypto trading contest finds LLMs can't trade crypto"

The operator's own site has never been loaded

nof1.ai returns HTTP 429 on direct fetch and has never been loaded by this project. Every figure above is secondary reporting, cross-checked between two independent sources that agree. Treat it as reported, not audited.

There is also a conflicting set of balances in this project's own earlier notes, which records different final figures for the same six models. Those are almost certainly a mid-competition snapshot: they were taken from a search aggregation rather than a fetched results page, and they disagree with the precise percentages on the losing side while agreeing on the ordering and on which two models finished up. The set shown above is the one that was verified by fetching the results summary and cross-checked against a separate aggregation, and it is the one used everywhere in this product.

One more thing is missing, and it matters here: no inference-cost, token-usage or dollar-cost-of-compute figure was ever published for Alpha Arena. So there is no external precedent for a return-per-credit number. Ours is the on-chain credit_spent counter and nothing else.

arXiv 2609.05663, production telemetry#

"What LLM Trading Agents Actually Do in Production: A Six-Month, Population-Scale Record from Two Fleets", DXRG (DX Research Group), submitted 2026-09-04.

Scale	Value
Single-model invocations	7.5M
On-chain actions	~300k
Multi-tool turns	231,638
Fills	14,596
Fleet one	DX Terminal Pro: 3,505 user-funded vaults trading real ETH in Base memecoin markets over 21 days, February to March 2026
Fleet two	a live alpha fleet of 500 to 599 user-created agents, 91 to 117 concurrently active, on Hyperliquid perps, June to August 2026

Findings, as reported#

Median leverage was 5.0x in every volatility sextile, across a 5.7x spread in realised volatility. The Spearman correlation between volatility and leverage was -0.001 at p = 0.92. Agents did not size down when volatility rose.
No directional edge versus benchmark. The alpha fleet's roundtrip win rate was 41% against a 50% benchmark.
The operating layer explained leverage better than the agent's own strategy text did: +0.425 leverage per risk-slider level, and agent fixed effects explained 60% of behavioural variance.
A leaderboard render boundary coincided with a 1.75x jump in position-taking right at the top-three cutoff.
43.2% of positions saw a favourable 300 bps or larger move within 24 hours, and 49.3% of those still closed at a loss. A mechanical stop or take-profit bracket would have recovered +39.0 bps per position.

What the preprint is worth#

Two caveats the paper does not hide either

It is a preprint. arXiv, submitted 2026-09-04, no peer review.

It is the vendor reporting telemetry on its own two products, not an independent audit of anybody else's agents. Directionally informative; not neutral third-party evidence.

On sourcing: the paper's existence and abstract were confirmed through a search summary and corroborating pages at huggingface.co/papers/2609.05663 and github.com/ProjectDXAI/continuous-record-llm-trading-agents. The arXiv HTML page itself was not loaded directly, so the exact author names are unconfirmed.

Why this is in the documentation#

Two of the paper's findings describe things Stags itself has, so they are read here as warnings rather than as trivia.

Finding	What Stags has	How it is read
Median 5.0x leverage regardless of volatility	a 5x hard cap, `MAX_LEVERAGE = 5`	A cap is a ceiling, not a strategy. The evidence says agents sit at the ceiling, so a 5x cap should be read as a 5x expectation.
A leaderboard boundary moved behaviour by 1.75x	return per credit, computable on chain and inherently rankable	A public ranking is an incentive to take more risk near its cutoff. Worth knowing before building the ranking.
Risk-slider level explained leverage better than strategy text	a risk slider that is the main control	The operating layer matters more than the prompt. That argues for the slider being conservative by default.

The specification's own conclusion, repeated here without softening: agent mode should be sold as what it is, and Passive mode is the honest default and likely the better product for most launches.

Claims this page will not make#

Not written anywhere	Why
"2.5 weeks"	The season ran 17 days. About two weeks.
"Models lose money", with no mention of the two that did not	Two of six finished up, and omitting that is a claim the reader can disprove in one search.
Citing the preprint as a study of the industry	It is a study of its own publisher's userbase.
Any return-per-credit benchmark drawn from outside this repository	None exists publicly. Alpha Arena published no compute cost at all.