Designing Agentic AI-Based Screening for Portfolio Investment

A team of researchers from leading institutions has published a groundbreaking paper on arXiv titled "Designing Agentic AI-Based Screening for Portfolio Investment." The work introduces a three-layer AI architecture specifically designed for quantitative portfolio management. At its core are two specialized large language model (LLM) agents: one screens companies based on desirable financial fundamentals, while a second agent analyzes news sentiment. These agents then deliberate to generate and agree upon buy and sell signals, dramatically narrowing the pool of candidate assets from a large portfolio.

This agentic screening process creates a unique theoretical framework where the final number of assets in the portfolio is itself a random variable, determined by the AI's screening. The researchers introduce the concept of 'sensible screening' and prove that, even with mild screening errors, the squared Sharpe ratio of the resulting portfolio consistently estimates its target. Empirically, the system was rigorously tested on S&P 500 data from 2020 to 2024. The results showed it achieved superior Sharpe ratios—a key measure of risk-adjusted return—when compared to both an unscreened baseline portfolio and conventional screening approaches. After screening, a high-dimensional precision matrix estimation procedure is applied to determine the optimal portfolio weights, completing the automated investment pipeline.