2026 · arXiv:2605.23007 · q-fin.TR
Can agentic loops optimize noisy rewards?
FunSearch and AlphaEvolve showed that an LLM in an evolutionary loop can rewrite code toward better solutions on deterministic targets — matrix multiplication, bin packing, compiler heuristics. Markets are the opposite: the reward is stochastic, non-stationary, and trivially easy to overfit. We put an LLM-driven evolutionary loop to that harder test — rewriting the code of a trading strategy, scoring each candidate against a held-out backtest, and keeping and recombining what survives — to ask whether an agentic loop is doing research or just p-hacking.
HOW IT WORKS
An LLM proposes a diff or a full rewrite to the parts of the strategy marked mutable; everything else — simulation, data, PnL accounting — stays fixed, so gains reflect real algorithmic change rather than evaluation artifacts. A structured population with quality-diversity selection and island migration keeps the search broad while it compounds on its best solutions, and mutation queries route across several frontier and efficient-tier models rather than betting on one.
WHAT WE FOUND
Gains that hold out of sample. Out-of-sample Sharpe improves by 0.6 to 1.8 points across runs; the best run reaches a test Sharpe of 5.65.
Not p-hacking. Against a conservative best-of-K null, the selected strategy sits roughly 45 standard deviations into the tail on held-out data — and the out-of-sample curve rises rather than decays, the opposite signature of overfitting.
Better forecasts, too. In the forecasting-only run, an evolved feature set roughly doubles out-of-sample R² over a three-feature baseline — though higher accuracy only becomes PnL once execution is re-tuned.
Model-agnostic. Best results trace their lineage through three to five distinct LLMs, so the engine compounds with each new model release rather than betting on one.
OPEN PROBLEMS
Metrics aren't PnL. Higher forecasting accuracy underperforms out-of-sample until execution hyperparameters are re-tuned.
Bigger search, more overfitting. The largest search space gives the best risk-adjusted result but the widest validation-to-test retention gap.
How far can it go? The objective isn't gradient-friendly, so there's no inner tuning loop yet — and the deeper question stands: how far can agentic loops push as reward functions get noisier?
SCOPE
Conditional on the backtest's market model — exchange-aggregated data that isn't directly tradable, full fill at the limit price, a specific fee and impact model. Results do not transfer out-of-the-box to live trading; live assessment is future work.
Kvasiuk, Li, Colegrove, Münchmeyer
University of Wisconsin–Madison & Event Horizon Labs
Read the paper →