From News to Signals: Agentic Workflows for Qualitative Investment Research
提出一种智能体工作流，将新闻分析分解为显式步骤，实现定性信息的规模化、透明化处理，并展示其提取的定性信号与传统因子互补，对量化投资研究者有方法论参考价值。
Investment research relies heavily on qualitative information embedded in narrative sources, yet prior text-based approaches typically compress interpretation into static features, obscuring the reasoning process by which narratives are assessed. This article introduces an agentic workflow that treats qualitative interpretation as a structured research process, decomposing news analysis into explicit stages such as relevance filtering, thematic identification, scoring, and aggregation. Using high-frequency firm-level news as a case study, the article makes four contributions. First, it presents a scalable agentic architecture that mirrors key elements of human qualitative research while operating at institutional scale. Second, it provides a structured analysis of how agent design choices affect signal quality, stability, cost, and interpretability, making trade-offs explicit. Third, it demonstrates how qualitative corporate factors can be extracted from news and mapped to economically meaningful constructs that complement traditional quantitative factors. Fourth, it presents empirical evidence that selected qualitative signals exhibit distinct long–short asymmetries, heterogeneous economic roles, and weak correlations with established factors. The primary contribution is methodological but critical to the future of investment success: By formalizing interpretation as an explicit and governable workflow, the framework enables qualitative judgment to be scaled with transparency and empirical discipline while preserving clarity about its economic role and limitations in systematic investment research.