Hubble Uses LLMs to Automate Alpha Factor Discovery

Hubble introduces an agentic, LLM-driven framework for automated alpha factor discovery that enforces deterministic safety and interpretability constraints. The paper reports experiments on a panel of 30 U.S. equities across 752 trading days, evaluating 181 syntactically valid factors from 122 unique candidates over 3 rounds and reaching a peak composite score of 0.827 with 100% computational stability.

Hubble constrains LLM generation using a domain-specific operator language and executes candidate formulas inside an AST-based sandbox to prevent unsafe operations and ensure deterministic evaluation. The system measures candidate factors with a statistical pipeline centered on:

• •RankIC (cross-sectional Rank Information Coefficient)
• •annualized Information Ratio
• •portfolio turnover

An evolutionary feedback mechanism returns top-performing factors and structured error diagnostics to the LLM, enabling iterative refinement across multiple generation rounds. The authors emphasize syntactic validation, computational stability checks, and a closed-loop that blends heuristic search with rigorous backtest hygiene.

Hubble directly addresses common failure modes of automated factor mining, including brittle, overfit formulas from genetic programming and opaque black-box generators. By pairing LLM creativity with a deterministic AST sandbox and explicit financial metrics, the framework trades unconstrained expressivity for reproducible, auditable discovery. This sits at the intersection of agentic LLM research and practical quant engineering: it leverages natural-language-driven search while preserving the kinds of guards quants require for model governance and backtest defensibility.

Limitations and caveats: The experiments use a modest universe and historical backtests; real-world performance will depend on transaction cost modeling, out-of-sample validation across regimes, and robustness to market microstructure effects. The LLM acts as a heuristic generator, not a guarantee of alpha, so human-in-the-loop validation and risk controls remain essential.

Replication at scale, formalized out-of-sample testing protocols, integration of transaction-cost-aware simulators, and potential open-source releases of the operator language and sandbox. If those follow, Hubble could become a practical template for safe, auditable LLM-assisted searching in quantitative research.