Research

A multi-agent LLM system called Agora reached a long-short Sharpe ratio of +1.87 on a sealed 91-day CSI 1000 holdout, beating its closest baseline's best-seed result of +1.33, but the authors flag their own headline number as a single, unrepeated run

Five specialised LLM roles jointly discover and refine trading factors across eight evolving skill libraries with information barriers to the holdout set. The baseline's seed variance tells the more important story: mean performance across three seeds dropped to -0.755, and Agora's own full-system seed stability remains uncharacterised.

· Source: arXiv (Panda AI)


The single most useful number in this paper isn't Agora's +1.87 Sharpe, it's what happened to its own comparison baseline when researchers actually checked robustness. AlphaGen-PPO hit +1.33 at its favourable seed, a headline-worthy result on its own, then collapsed to a mean of -0.755 across three seeds. That's not a small variance band, it's the difference between a strategy worth deploying and one that would have blown up an account, produced entirely by which random initialisation the training run happened to draw.

Agora's authors deserve credit for stating plainly that their own full 100-round system ran once, and that its seed variance remains uncharacterised. That's an unusual level of candour for a paper built around a strong headline metric, and it should be read as the load-bearing caveat rather than a footnote, because the AlphaGen comparison sitting right next to it demonstrates exactly how much a single favourable seed can flatter an otherwise unstable system. Until Agora's own multi-seed numbers exist, the +1.87 figure is a data point, not a demonstrated capability.

The long-short structure is the second caveat worth taking seriously. The bottom decile of Agora's signal produced -27.39% annualised, the top decile only +6.46%, meaning the entire positive Sharpe depends on the shorting leg working cleanly. The authors are explicit that this makes the signal unsuitable for long-only investing, which is a meaningful constraint given how much retail and even institutional deployment of AI trading signals defaults to long-only implementation for cost, borrow-availability, and mandate reasons that have nothing to do with the model's actual edge.

Read alongside the growing pile of LLM-trading-agent papers this year, Agora's real contribution may be methodological rather than performative: information barriers between the factor-discovery agents and the sealed holdout set are a genuine attempt at preventing the kind of look-ahead leakage that has quietly inflated backtest results across the field. Whether the underlying factor-discovery approach survives a proper multi-seed stress test, the same test that broke its own baseline, is the open question the next version of this research needs to answer before anyone treats the CSI 1000 result as more than a promising single draw.


Read the original: arXiv (Panda AI) - A multi-agent LLM system called Agora reached a long-short Sharpe ratio of +1.87 on a sealed 91-day CSI 1000 holdout, beating its closest baseline's best-seed result of +1.33, but the authors flag their own headline number as a single, unrepeated run. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.