AI Trading Agents Achieve Emergent Alpha Through Self-Evolution

Traditional automated alpha mining in finance often fixes the scoring function while varying the search algorithm, leading to overfitting. This new research proposes Sealed Joint Search (SJS), a framework designed to allow LLM-based trading agents to jointly evolve both alpha factors (trading signals) and their evaluation criteria (scoring functions). SJS includes structural conditions to prevent self-confirmation bias, ensuring robust discovery. The Agora system, an empirical implementation of SJS, features five LLM agent classes communicating through distinct channels, evolving eight skill libraries and alpha libraries. Crucially, three evaluators generate reports, carrying forward disagreements rather than forcing a consensus, which helps maintain diverse perspectives. Tested on the CSI 1000 index with a 91-day sealed holdout, Agora achieved an impressive Sharpe ratio of +1.87, significantly surpassing the best baseline of +1.334. The study highlights that the key metrics driving this performance emerged organically through the agents' self-evolution, rather than being pre-designed. This demonstrates a powerful new paradigm for AI-driven financial strategy development.