Hubble: An LLM-Driven Agentic Framework for Safe and Automated Alpha Factor Discovery

Researchers introduce Hubble, an LLM-driven framework that automates alpha factor discovery in quantitative finance by using large language models constrained by safety mechanisms to generate and refine predictive trading factors. The system achieved a composite score of 0.827 across 181 evaluated factors on U.S. equities, demonstrating that combining AI-driven generation with deterministic safety constraints enables interpretable and reproducible factor discovery.

Hubble addresses a longstanding challenge in quantitative finance: discovering novel predictive factors amid massive search spaces and noisy financial data. Traditional genetic programming approaches often produce opaque, overfit models that fail in live trading. This research demonstrates a practical alternative by positioning LLMs as intelligent explorers within a constrained sandbox environment, guided by domain-specific rules and an Abstract Syntax Tree execution layer that prevents invalid or unsafe operations.

The significance lies in bridging interpretability with automation. Quantitative researchers have historically chosen between manual factor engineering—labor-intensive but interpretable—and algorithmic generation—scalable but often uninterpretable. Hubble creates a middle ground by leveraging LLM reasoning capabilities while maintaining strict statistical and syntactic validation. The evolutionary feedback loop, which returns performance metrics and error diagnostics to the model, creates a closed-loop refinement process that mirrors human research iteration at machine speed.

For the quantitative finance industry, this methodology could democratize alpha discovery. Smaller hedge funds and asset managers without deep specialized talent pools could deploy such systems to compete with larger firms. The emphasis on 100% computational stability and the rigorous evaluation pipeline (RankIC, Information Ratio, turnover analysis) suggest the authors prioritize production readiness over academic novelty.

Looking ahead, the approach raises questions about factor durability in live markets. The 752-day backtest period is substantial, yet factor decay remains a critical concern in quantitative finance. Future developments may focus on incorporating market regime detection, robustness testing across asset classes, and scaling to larger universes beyond 30 equities. Regulatory scrutiny around AI-driven trading strategies will also shape adoption.

• →Hubble combines LLM intelligence with deterministic safety constraints to automate interpretable alpha factor discovery in quantitative finance.
• →The framework achieved a peak composite score of 0.827 across 181 evaluated factors with zero computational failures.
• →Evolutionary feedback mechanisms enable iterative refinement, reducing typical overfitting issues seen in genetic programming approaches.
• →The system bridges the gap between manual, interpretable research and fully automated but opaque algorithmic generation.
• →Results suggest LLM-driven factor discovery could democratize quantitative research access for mid-sized asset managers.