LLM Trading Agents in Production: Operating Layer Dominates, No Directional Edge

A six-month population-scale record from two live LLM trading fleets reports that risk and posture controls, agent fixed effects, and a leaderboard render boundary shape leverage and selection more than strategy text; leverage is volatility-blind, favorable excursions are rarely converted, and neither fleet shows a directional edge. The authors report robustness to day-clustered inference, permutation nulls, and a common-fee restatement.

Hugging Face Daily Papers — What LLM Trading Agents Actually Do in Production: A Six-Month, Population-Scale Record from Two Fleets

September 10, 20261 source76% confidence

80Strategic

Strategic score

Evidence strength72

Technical novelty80

Commercial potential70

This is not another model leaderboard. It is a six-month, population-scale production record from two live LLM trading fleets: DX Terminal Pro, with 3,505 user-funded vaults trading real ETH in Base memecoin markets over 21 days in February–March 2026, and the DXAP live alpha fleet, with 500–599 user-created agents (91–117 concurrently active) trading Hyperliquid perpetuals in June–August 2026. The record covers roughly 7.5M single-model invocations, about 300K onchain actions, 231,638 multi-tool turns, and 14,596 fills.¹

Context

The paper's central contrast is between strategy text and the operating layer. The measured levers are a risk slider, a posture slider, agent fixed effects, and a leaderboard render boundary. That contrast is what makes the record new relative to decision-quality evaluations alone: it treats the user interface, fleet lifecycle, and ranking display as first-class determinants of realized trading behavior.¹

Assessment

The strongest single identification is the leaderboard render boundary: a regression discontinuity reports 1.75x selection at the top-3 cut. The risk-slider coefficient (+0.425 leverage per level) and agent fixed effects absorbing 60% of variance are decompositional and observational, not causal. The volatility-blind sizing result—median 5.0x leverage in every volatility sextile—and the concentration result—11% of the book in one posture-slider cell holding 62% of liquidations—are descriptive but economically important. The upside-capture gap (43.2% of positions reach +300 bps in 24h; 49.3% of those close negative) and the +39.0 bps mechanical bracket are the clearest actionable findings.¹

Caveats

What would change the conclusion: audited, independent replication of the trade logs; pre-registered analysis of the operating-layer parameters; out-of-sample performance in a different market regime; or evidence that the 41% vs 50% roundtrip win-rate gap against a matched Hyperliquid retail benchmark disappears under alternative matching, fee, or funding assumptions. The paired-replay league's 'statistically indistinguishable' decision quality across frontier models on 416 scenarios may reflect low power or a horizon too short to separate models. The abstract does not report uncertainty for every headline, and user-funded vaults are not independent agents.¹

Source text

Hugging Face Daily Papers — What LLM Trading Agents Actually Do in Production: A Six-Month, Population-Scale Record from Two Fleets · September 4, 2026

Computer Science > Artificial Intelligence

Title:What LLM Trading Agents Actually Do in Production: A Six-Month, Population-Scale Record from Two Fleets

Abstract:We present a continuous, population-scale measurement record of autonomous language-model trading agents operating in production across two systems with one design lineage: DX Terminal Pro (3,505 user-funded vaults trading real ETH in Base memecoin markets for 21 days, February to March 2026) and the DXAP live alpha fleet (500 to 599 user-created agents all-history, 91 to 117 concurrently active, trading Hyperliquid perpetuals, June to August 2026). The record spans roughly six months, 7.5M single-model invocations with about 300K onchain actions, and a further 231,638 multi-tool turns producing 14,596 fills. Four findings carry the paper. First, the operating layer determines behavior more than anything written in strategy text: a risk slider explains leverage (+0.425 per level), agent fixed effects absorb 60% of variance, and a leaderboard render boundary causally routes selection (regression discontinuity 1.75x at the top-3 cut). Second, sizing is volatility-blind: median leverage is 5.0x in every volatility sextile, and one posture-slider cell (11% of the book) holds 62% of liquidations. Third, agents capture almost none of the upside they reach: 43.2% of positions saw at least +300 bps of favorable excursion within 24h, yet 49.3% of those closed with a negative trade return; a mechanical bracket recovers +39.0 bps per position. Fourth, neither fleet shows a directional edge. The DXAP fleet is not profitable and trails a matched Hyperliquid retail benchmark (41% vs. 50% roundtrip win rate). A paired-replay league of frontier models on 416 captured production scenarios finds decision quality statistically indistinguishable at this horizon, while choice stability differs sharply across model families. Every headline survives day-clustered inference, permutation nulls, and a common-fee restatement; the paper closes with a 17-rule methodology canon bought with our own retractions.

Access Paper:

Current browse context:

Bookmark

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

The phrase 'operating layer determines behavior more than anything written in strategy text' bundles three different designs. The risk-slider coefficient is an observational regression; agent fixed effects absorb variance without identifying a causal channel; only the leaderboard discontinuity is quasi-experimental. The causal claim is strongest for selection at the top-3 cut, weaker for leverage and liquidation. The abstract's unification of these effects should not be read as three independent causal estimates.

Median leverage 5.0x in every volatility sextile is a null on volatility targeting. Its force depends on unstated choices: realized versus implied volatility, estimation window, whether sextiles are cross-sectional or time-series, and whether leverage is capped by the same risk-slider range. If the leverage control is discrete and range-bound, a flat median could be mechanical. The full methodology canon is needed to distinguish a behavioral anomaly from a control artifact.

The 49.3% negative-close figure is conditional on the 43.2% of positions that first reached +300 bps within 24h. It is not a general negative-trade rate. The economic loss is failure to convert reached upside, and the +39.0 bps bracket recovery is a per-position gross estimate that may not survive changes in fill rates, fees, or adverse selection. The abstract reports both numbers, but the compression invites misuse.

The 41% vs 50% roundtrip win-rate comparison against a 'matched Hyperliquid retail benchmark' depends entirely on the matching. Market, period, leverage, size, and fee/funding treatment are not specified in the abstract. A nine-point gap is meaningful only if the benchmark is credible; otherwise it may reflect benchmark construction. 'Not profitable' also does not by itself establish negative net expectancy after all costs.

'Statistically indistinguishable' decision quality across frontier models on 416 captured production scenarios is not evidence of equivalence. It can arise from low power, scenario selection that compresses variance, or a horizon too short for model differences to appear. The separate finding that choice stability differs sharply across model families is about consistency, not accuracy, and should not be merged with the accuracy null.

The 17-rule methodology canon 'bought with our own retractions' signals methodological reflexivity, not validity of the four headline findings. Retractions may improve rules, but if the final canon was assembled after the headline analyses, it cannot be treated as pre-registration. Readers should ask whether the 17 rules were fixed before the reported tests or adapted post hoc.

The source text above remains the property of its publisher and is reproduced for commentary.

Hugging Face Daily Papers — What LLM Trading Agents Actually Do in Production: A Six-Month, Population-Scale Record from Two Fleets

September 4, 2026endistinct angle

What LLM Trading Agents Actually Do in Production: A Six-Month, Population-Scale Record from Two Fleets

↑ Back to top