Designing Agentic AI-Based Screening for Portfolio Investment

Battle Count: 138

Directly addresses portfolio construction and stock screening using AI. The two-stage pipeline (screening + weighting) mirrors real-world quant fund workflows. Empirical results on S&P 500 with Sharpe ratios of 1.09 (medium-term) and 8.16 (short-term) are highly relevant. The paper provides concrete implementation details including LLM prompts, factor selection, and precision matrix estimation methods. The finding that human analyst integration degrades AI performance is practically significant for quant fund design.

Requires: (1) LLM API access (ChatGPT-3.5/4o) with custom prompt engineering for the LLM-S agent; (2) FinBERT fine-tuned model for sentiment analysis; (3) Multiple high-dimensional precision matrix estimation methods (nodewise regression, residual nodewise, POET, deep learning, NLS); (4) Portfolio optimization solvers for GMV, MV, MSR; (5) Rolling window backtesting infrastructure; (6) Data pipelines for CRSP/Compustat fundamentals, I/B/E/S analyst data, and news articles. The intersection/union consensus logic and method-objective selection add further complexity.

The paper uses specific LLM versions (ChatGPT-3.5, ChatGPT-4o) with defined knowledge cutoffs to prevent look-ahead bias. However, LLM outputs can vary across runs (authors note they reran the Agentic AI screen twice with similar results). Data sources (CRSP, Compustat, I/B/E/S, HuggingFace news) are standard but require subscriptions. The quantitative methods (nodewise regression, POET, NLS) are well-established. No code repository is mentioned.