International Edition
Latest News
Business

LLM Agents Deliver 34.5% Return in Study, Outperforming SPY With Higher Drawdown

Market strategies using automated intelligence generated a 34.5% return from March 11 to August 28, 2026, according to a paper-trading account study shared on arXiv. That performance surpassed the 14.4% gain delivered by the S&P 500 fund SPY…

LLM Agents Deliver 34.5% Return in Study, Outperforming SPY With Higher Drawdown

Market strategies using automated intelligence generated a 34.5% return from March 11 to August 28, 2026, according to a paper-trading account study shared on arXiv.

That performance surpassed the 14.4% gain delivered by the S&P 500 fund SPY over the identical period.

The broker-recorded paper account achieved this advantage with substantially higher maximum drawdown than the benchmark index, as detailed in the arXiv preprint. Researchers Renee Jia, Di Mu, Yao Yu, and Tianrui Liu evaluated the hybrid multi-agent trading system, which pairs large language model analysts with momentum-based portfolio construction.

Company-news and macroeconomic agents process incoming data to return structured reports to a central coordinator. This coordinator sets stock weights and cash reserves based on quantitative scoring.

The system incorporates technical indicators, macro trends, and sentiment scores into its allocation engine. A production score combines these analytical inputs before a momentum module selects the top candidates for the portfolio.

The macro score calculates as half the quantitative macro value plus half the model macro value, rounded to one decimal place. The momentum module selects the top 10 candidates based on the current book, operating alongside a later rule using a 15-name rank buffer.

The live paper account operated using deterministic fallbacks for trade execution. Researchers examined LLM-enabled configurations through retrospective replays to test different decision-making speeds.

Reconstructed policies finished above matched SPY and QQQ references whether assessments occurred in the same session or deferred to the next. Design tests showed that a constant-cash control delivers return estimates close to stance-based sizing.

Shorter report contracts improve observed parsing speed for the models but can inadvertently change their stock recommendations. The reliability paradox means that making LLM outputs easier for downstream code to parse alters the underlying economic guidance.

Capital allocation rules cap single positions. The framework adjusts cash reserves based on macroeconomic scores.

Rank values derive from descending rank values assigned to selected assets with a high of 3.0 and a low of 0.2, combined with score values calculated as the production score divided by 100 squared.

Raw allocations mix 95% of the rank value with 5% of the score value, while cash reserves quantize macro values against 45, 60, and 75 thresholds. Target allocations normalize to invested capital after subtracting cash reserves, and the system caps positions while redistributing excess with a maximum position percentage of 25.

About the author: Business Desk

Business Desk at Archynewsy.