Research

A new arXiv paper describes AQuA, a pair of AI research agents that recursively improve their own quant strategies, reporting a 2.50 Sharpe equity strategy that stayed positive every year from 2021 through 2025

Unlike most of the trading-agent literature this site has covered, AQuA isn't an agent that trades, it's an agent that does the research work of building trading strategies, then uses its own prior results to get better at that job.

· Source: arXiv (Guo, Huang, Gao, Li, Ge, Kuang, Wang)


AQuA, published on arXiv on 13 August, splits the work of quant research into two AI-driven components: one hunting for symbolic factors, the other building trainable models, both able to draw on the outcomes of their own earlier experiments to sharpen the next round. The headline numbers are strong for a system this early in its life. The factor-discovery side hit an information coefficient of roughly 0.190 in crypto markets, and the model-building side produced a US equity strategy with a Sharpe ratio up to 2.50 that stayed positive in every year from 2021 to 2025, a run that spans two very different regimes for factor-based investing.

That result sits in interesting contrast to most of the LLM-trading-agent research this site has tracked, which has generally been a story of fragility rather than robustness. The Backtrader-Bench paper found tool-augmented agents scoring 90 percent on standard benchmark questions and near-random on harder, less-curated ones. The 77-study reproducibility review and the Yao/Zheng execution-assumptions audit both found that headline returns in this literature often quietly assume away real-world frictions like slippage and latency. AQuA's framing sidesteps some of that criticism by not claiming to be a live trading agent at all, it's positioned as a research tool that proposes strategies for humans or downstream systems to evaluate, which is a narrower and more defensible claim than "an AI agent that trades profitably."

The multi-year consistency of the 2.50 Sharpe result is the detail worth sitting with. A lot of AI-discovered strategies in this space look impressive on a single backtest window and fall apart the moment the sample period changes, which is exactly the overfitting risk the reproducibility crowd has been flagging. Five years of positive returns, spanning the 2022 rate-hiking cycle and the 2023 to 2025 AI-equity run-up, is a genuinely harder bar to clear than most single-period claims in this literature, even allowing for the usual caveat that backtested Sharpe ratios tend to shrink once a strategy meets live capital.

Where this fits into the broader picture: recursive self-improvement in a narrow, bounded domain like factor research is a much more tractable claim than recursive self-improvement in general intelligence, and it's a plausible near-term use case for agentic AI inside quant shops that already run large factor libraries. Worth watching whether AQuA's approach gets adopted or replicated by a real allocator rather than staying a research-paper result, since that's the point where claims like this usually get stress-tested against actual capital and actual slippage.


Read the original: arXiv (Guo, Huang, Gao, Li, Ge, Kuang, Wang) - A new arXiv paper describes AQuA, a pair of AI research agents that recursively improve their own quant strategies, reporting a 2.50 Sharpe equity strategy that stayed positive every year from 2021 through 2025. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.