From Hypotheses to Factors: Constrained LLM Agents in Cryptocurrency Markets

The Thesis

This paper shows that large language model (LLM) agents — AI systems that can reason, propose ideas, and execute code — can discover useful trading signals in cryptocurrency markets when their search is tightly constrained by a formal protocol. The key insight is that unconstrained AI search tends to overfit: it finds patterns that look good in past data but fail in the real world. The authors address this by forcing the agent to work inside a domain-specific language (DSL, a restricted coding environment that limits what kinds of signals can be expressed), fixed data splits, and automatic transaction-cost penalties. The resulting portfolio, built only on 2020–2022 data, posted a 44.55% annualized return and a Sharpe ratio of 1.55 (a risk-adjusted performance measure where above 1.0 is generally considered good) over a 2024–2026 test window. The catch: cryptocurrency markets are volatile, lightly regulated, and prone to structural breaks, so results that look clean on paper may not translate to live trading.

Catalyst

LLMs capable of writing and debugging financial code became widely available only in 2023–2024, making automated hypothesis-and-test loops over structured financial data newly feasible. At the same time, the crypto market's rich, high-frequency public data makes it an ideal testbed — more tractable than equities, where survivorship bias and data licensing create additional hurdles. The maturation of point-in-time data infrastructure (data stored exactly as it was known at each historical moment, preventing accidental look-ahead) gave researchers a clean sandbox to test agent-driven factor discovery without the most common sources of backtesting fraud.

What's New

Earlier automated factor discovery systems — such as Alpha Mining pipelines and genetic programming approaches — searched over predefined mathematical operations without any natural-language reasoning layer. Those systems could combine existing signals but couldn't read research literature, form English-language hypotheses, or explain why a factor might work. This paper adds an LLM reasoning layer on top of a deterministic execution engine, so the agent proposes human-readable hypotheses first and then maps them to executable code, making both successes and failures auditable in a way that pure code-evolution approaches are not.

The Counter

A 44% annualized return in crypto is impressive on paper, but crypto markets routinely produce such numbers by chance — the asset class saw 100%+ drawdowns and recoveries in the same 2020–2026 window, so a lucky factor tilt toward beta (overall market exposure) could explain most of the result. The 2024–2026 out-of-sample window is also only about two years long, which is far too short to distinguish skill from luck in a market with as much regime variance as cryptocurrency. The paper does not report how many factor hypotheses the agent generated and discarded before finding the ones that made it into the final portfolio — if hundreds of hypotheses were tested and only winners reported, the 'out-of-sample' period may be contaminated by implicit selection pressure. Finally, the specific DSL and protocol design choices were made by the researchers, meaning the framework itself was tuned on knowledge of what kinds of crypto factors have historically worked — a subtler form of overfitting that no fixed data split fully prevents.

Resources

• https://arxiv.org/abs/2604.26747
• https://arxiv.org/abs/2306.11644
• https://en.wikipedia.org/wiki/Sharpe_ratio
• https://en.wikipedia.org/wiki/Factor_investing
• https://arxiv.org/abs/2409.03517

The Paper