Beyond Prompting: Autonomous Factor Investing via Agentic AI

Battle Count: 72

This paper is directly and highly relevant to quantitative trading. It presents a complete end-to-end framework for autonomous factor discovery and portfolio construction, with rigorous out-of-sample validation on U.S. equities. The reported Sharpe ratio of 2.75 and annualized return of 54.81% (gross) / 305.57% cumulative net of costs are directly actionable for practitioners. The framework addresses core practitioner concerns: data mining, overfitting, transaction costs, turnover constraints, and economic interpretability. The comparison against traditional AI frameworks demonstrates clear incremental value. The methodology is designed for institutional deployment with explicit attention to implementation feasibility.

The framework involves multiple interconnected components: LLM-based agent with memory and reflection capabilities, symbolic regression engine, deterministic execution layer, multi-metric evaluation pipeline, gatekeeping logic, and LightGBM aggregation. Requires access to CRSP data, LLM API or local deployment, and substantial computational resources for iterative backtesting. The closed-loop architecture with state management, policy updates, and structured logging adds engineering complexity. However, the modular design and formalized algorithm (Algorithm 1) provide a clear implementation roadmap. The factor grammar and operator set are well-defined, reducing ambiguity in implementation.

The paper provides a project homepage (https://allenh16.github.io/agentic-factor-investing/) with interactive results, methodological documentation, and replication details. The data pipeline is described as deterministic and fully auditable. However, the specific LLM model used is not explicitly named, and the exact hyperparameters for the agent's search policy are not fully detailed. The CRSP data is standard but requires purchase. The closed-loop workflow is formally specified with Algorithm 1, and all evaluation protocols are pre-committed.