A team of researchers put large language model agents inside an end-to-end equity trading system — connecting market analysis, portfolio allocation and execution — and the broker-recorded paper account returned 34.5% from March 11 to August 28, 2026, against 14.4% for the S&P 500 fund SPY, but with a substantially larger maximum drawdown, according to the arXiv preprint the team published on November 8.
Hugging Face, which carried the project under the title Inside the Trading Loop: Markets, Models, and Moves with LLM Agents, frames the question bluntly: what happens when AI agents go beyond analyzing markets and start influencing real portfolio decisions? The answer, the sources show, is a mixture of outperformance, unexamined risk and a finding that should unsettle anyone deploying these systems in production.
The study, authored by Renee Jia, Di Mu, Yao Yu and Tianrui Liu and posted to arXiv before peer review, describes a hybrid architecture. LLM analysts handle company news and macroeconomic conditions, returning structured reports to a coordinator that sets stock weights and cash reserves.
Momentum-based portfolio construction and rule-based execution sit underneath. The LLM does not press the buy button; it informs the weights above a deterministic execution layer, per the paper.
Broker paper account returned 34.5% versus 14.4% for SPY
- The broker-recorded paper account returned 34.5% versus 14.4% for SPY over identical dates — but the raw advantage came with a noticeably larger maximum drawdown, per the paper.
- The recorded account ran on deterministic fallbacks; the LLM-enabled configurations were examined only through retrospective replays, a distinction the authors themselves flag.
- Shorter report contracts improved parsing but changed the agents’ recommendations — an effect the team calls a reliability paradox.
Here is the math behind the headline number. The comparison window runs March 11 through August 28, 2026, the same dates for both the account and the SPY benchmark, as ReadSearch summarizes of the preprint. A 20-percentage-point gap in five and a half months is large by any institutional standard.
But the team’s own materials state the account used deterministic fallbacks while LLM-enabled strategies were studied through replays, meaning the live, recorded performance was not generated by the agents themselves. That single distinction is where most of the marketing noise around this paper would collapse.

| Configuration | Return (Mar 11–Aug 28, 2026) | Max Drawdown | Recorded or Replayed |
|---|---|---|---|
| Broker-recorded paper account (deterministic fallbacks) | 34.5% | Substantially higher than SPY (per study) | Recorded |
| SPY benchmark | 14.4% | Lower (per study) | Benchmark |
| Reconstructed LLM policies | Above matched SPY and QQQ references | Not stated in abstract | Replayed |
But the balance sheet tells a different story, and the paper does not hide it. Higher returns arrived with a larger maximum drawdown — the depth of peak-to-trough loss an investor would have had to sit through. In quantitative finance, that trade-off is the entire appraisal. A 34.5% return financed by deeper drawdowns reflects a portfolio taking more risk, not necessarily a model making better bets.
Baseline of a constant-cash control
One of the paper’s sharper design comparisons involves cash authority. A constant-cash control — a rule that simply holds a fixed cash level — produced a return estimate close to the agents’ stance-based sizing, per the study’s abstract.
In plain terms: a large share of the performance the system attributes to agent intelligence can be approximated by a rule that ignores the agents entirely. That is not a fatal result for LLM-based trading. It is, however, a results-bar the system’s next iteration has to clear before anyone calls the intelligence additive.
The coordinator’s cash levels were quantized into three stances, and the reconstructed portfolio engine caps any single position at 25% and draws from a top-10 momentum list with a 15-name rank buffer, according to the technical detail carried on the arXiv page for the preprint.
Stance-based cash plus a momentum spine means the system’s engine room is recognizable to any quant desk. The question the paper poses is when the LLM layer changes outcomes — and when it merely burns computation.
When parsing changes the recommendation
The reliability paradox is the paper’s most transferable finding for anyone building AI systems beyond finance. The team reports that shorter report contracts — Structured LLM outputs that were easier for downstream components to parse — improved observed parsing but, unexpectedly, changed the agents’ recommendations. Compression of a report is not neutral; it alters what the coordinator sees, and therefore what the coordinator thinks.
That result sits alongside two other axes the authors probe: selective invocation, examining when invoking agents on demand can preserve decision impact while cutting unnecessary computation, and agent authority — how reasoning translates into execution, risk exposure and stateful portfolio management.
The abstract states that reconstructed policies finished above matched SPY and QQQ references when AI assessments were used in the same session or deferred to the next session, which the team reads as evidence the pipeline works consistently regardless of timing choices.
For the broader market, the relevance is practical. Asset managers across the board are pushing AI from analytics into allocation — the same intersection this paper tests.
The next step the sources leave open is the one that matters: none of the three outlets report live, broker-recorded results for the LLM-enabled configuration itself. Until someone runs those agents with their hands on the account motion by motion, the 34.5% belongs to the deterministic fallbacks — and the agents owe the market a verified quarter.
Disclaimer: The information provided in this article is for educational and informational purposes only and does not constitute financial advice.
Related reading
- Scattered Spider Hacker Gets Six Years for Crypto Theft Scheme
- Nvidia Raises Fiscal 2028 Forecast, Keeps Huang Supporters at Bay
- Alyssa Farah Griffin Compares System Failures to Eerie Feeling Before COVID Pandemic (world-today-news.com)
- Jessie J Returns to Instagram After Breast Reconstruction Surgery (newsdirectory3.com)