Graphstrat: An LLM-to-Econometrics Pipeline Using Graph-Structured Relational Analytics Measurements of Corporate Disclosures

Abstract

GraphSTRAT (Graph-Structured Relational Analytics for Theory), a Graph-based Retrieval-Augmented Generation (GraphRAG)-based Large Language Model (LLM) method, leverages domain-specific ontology extraction to build context-aware measures for text-based business research, creating: a novel, integrated Dynamic Capabilities Index (DCI); an artificial intelligence (AI) deployment index (AIDI); and a Dynamic Capabilities Knowledge Graph Explorer for a corpus of corporate disclosure filings. The first paper demonstrates a new method for extracting actor–action–object relationships from 10-K filings by operationalizing Dynamic Capabilities (DC) theory and shows that the resulting Dynamic Capabilities Knowledge Graph Explorer provides provenance and visibility, allowing researchers to inspect and trace claims to exact passages. The second paper simultaneously evaluates DC and the maturity of AI deployment, creating DCI and AIDI from quantifiable data to link DC and AI to a firm’s performance. The results indicate that AIDI predicts near-term profitability pressures and increases in R&D intensity, and that DCI predicts market-based outcomes and strategic reconfiguration proxies. A directional predictive relationship from AIDI to later DCI is also observed, consistent with AI implementation inducing governance, routines, and reconfiguration capabilities. The results demonstrate how relational, auditable measurement improves theory alignment and enables new tests of technology-capability complementarities in large-scale archival settings. The research extends text-as-data methods from word-level themes to theory-anchored relational structures; introduces auditable LLM-based construct measurement that enables relational-level validation; and provides new empirical evidence and insights into large firms’ Dynamic Capabilities and AI deployment maturity.