A new arXiv paper reports lifting the average Sharpe ratio of seven trading strategies from -0.298 to 0.538 using large language models [S1]. The best-performing strategy improved 199% on a relative basis [S1]. The same paper admits that letting LLMs rewrite trading code directly causes hallucinated edits, strategy drift, and backtest overfitting [S1]. How do you stop a model from confidently inventing its own financial reality?

Why naive LLM trading fails

Quantitative strategy optimization is still largely manual work. Domain experts hunt for weak signals, tune risk-control rules, and validate each revision through repeated backtesting [S1]. It is slow, expensive, and human-bound.

You might think an LLM could automate this. Feed it the strategy code, let it find the bugs, spit out improvements. The EVOQUANT authors, a team from HKU led by Jie Mao and Changlun Li [P2], tried exactly that and found three failure modes. LLMs hallucinate edits, inventing code changes that look plausible but break the strategy's logic. They cause strategy drift, pushing the system away from its original intent. And they overfit to backtests, producing strategies that look great on historical data and fail in live markets [S1].

The verifier pipeline

EVOQUANT's answer is a multi-stage verification pipeline. The framework uses LLMs to diagnose performance bottlenecks in a trading strategy, then generates candidate edits with semantic control, meaning each change stays within the strategy's intended logic rather than wandering into new territory [S1]. A verifier then screens the candidates through multiple checks before any edit is accepted.

The framework also distills optimization experience into reusable knowledge, so each round of improvement feeds the next [S1]. The system learns from its own fixes, building a library of what works and what does not.

The numbers

The team tested EVOQUANT on seven representative strategies: four from the A-share market (Chinese equities) and three from cryptocurrency markets [S1]. The average test Sharpe ratio, a measure of risk-adjusted return, rose from -0.298 to 0.538 across all tested strategies [S1]. The best-performing strategy achieved a 199% relative improvement [S1].

Ablation studies and stress tests under stricter conditions were conducted to validate the framework's effectiveness [S1].

What it means

A Sharpe ratio of -0.298 means a strategy is losing money with high volatility relative to its returns. A reading of 0.538 means it is generating positive returns with better risk control. The swing from negative to positive is the difference between a fund shutting down and a fund surviving.

The key innovation is not the LLM itself. It is the verifier. Raw language models are unreliable code editors in finance because they optimise for plausibility, not profitability. By inserting a multi-stage verification pipeline between the model's suggestions and the live strategy, EVOQUANT tries to catch hallucinated edits and overfitting before they reach production.

The knowledge distillation component matters too. Most automated strategy tools start from scratch each time. EVOQUANT accumulates experience, so the hundredth optimisation round draws on lessons from the first ninety-nine. Whether that accumulated knowledge generalises across market regimes is an open question.

What it means for business

For a two-person quant shop running a handful of strategies on Chinese equities or crypto, the promise is clear: less manual tuning, faster iteration, and a system that gets smarter over time. Today, finding and fixing a weak signal in a trading strategy means days of expert analysis and repeated backtesting. EVOQUANT's framework aims to compress that into an automated loop.

The cost implications are real. Quantitative analysts are expensive, and strategy optimisation is their most time-consuming task. A framework that can diagnose bottlenecks and propose controlled edits could let a small team manage more strategies with the same headcount.

But the operational risks are equally real. The paper itself acknowledges that LLM-generated strategies carry inherent risks of hallucination and overfitting [S1]. The verifier pipeline is the authors' proposed mitigation, not an independently validated guarantee. A fund that deploys this without its own risk controls on top is trusting a preprint.

The seven-strategy test covers A-shares and crypto only [S1]. A firm trading US equities, FX, or commodities has no evidence this works in those markets. And the abstract does not mention live trading slippage, execution costs, or market impact, all of which can erode backtested returns.

What we don't know yet

The paper is an arXiv preprint, not peer-reviewed [S1]. All performance metrics are self-reported by the authors. No independent replication or audit of the Sharpe ratio improvements has been published.

The framework has not been deployed in live production trading, so the gap between backtested and realised performance is unknown. Slippage, execution costs, and market impact could materially change the results.

The test covers seven strategies across two markets. Whether EVOQUANT generalises to other asset classes, longer time horizons, or different market conditions is untested.

The authors' institutional affiliations beyond HKU are not fully specified in the available evidence [P2], and the paper does not appear to have a companion open-source code release linked from the arXiv entry.

The next concrete events to watch: whether the authors release code, whether the paper passes peer review, and whether any fund reports independent replication of the results.

If you want this kind of plain-English decode of every AI research paper that actually changes how businesses operate, the subscribe button is below.

Sources

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.