AlphaAgent: LLM-Driven Alpha Mining with Regularized Exploration to Counteract Alpha Decay

AlphaAgent splits factor mining across three LLM agents: an Idea Agent that proposes market hypotheses, a Factor Agent that turns them into regularized factor expressions, and an Eval Agent that validates and backtests them and feeds results back. The paper’s argument is that unconstrained LLM search produces factors that decay quickly, so exploration is bounded by three regularizers — originality enforced through an abstract-syntax-tree similarity measure against existing alphas, hypothesis-factor alignment judged semantically by an LLM, and complexity control through AST-based structural constraints. Published at KDD ’25 (pp. 2813–2822, doi:10.1145/3711896.3736838).

The implementation is derived from Microsoft RD-Agent, which the README credits explicitly, and backtests through Qlib on China A-share data pulled from baostock. Over CSI 500 and S&P 500 for January 2021 to December 2024 the authors report average annual excess returns of 11.0% (IR 1.5) and 8.74% (IR 1.05) after transaction costs; these are the authors’ own backtests, with no third-party replication published.

The repository needs care. The paper’s code is on the legacy-main branch, last committed in June 2025 — on 1 July 2026 the owner re-initialised the default main branch as a completely different project, a Tushare-based multi-factor framework whose reference dataset is distributed by Baidu Netdisk and which is marked “research use only” with no LICENSE file at all. Only legacy-main carries a license, MIT, and even there the copyright line reads “Copyright (c) Microsoft Corporation”, inherited unchanged from the RD-Agent fork rather than naming the paper’s authors.