Identifying episodes of fiscal austerity: An LLM-based approach

This paper introduces a hierarchical Large Language Model (LLM) framework for the automated identification of narrative fiscal shocks. We develop a multi-stage architecture to extract austerity episodes from IMF Article IV reports (2004–2020) for 17 OECD countries. Relative to manual coding, our approach improves replicability and auditability by generating a documented sequence of classification steps. Benchmarking against Adler et al. (2024), we find that the LLM-based classification aligns closely with the narrative benchmark, while differing on a small subset of episodes. Local projection estimates indicate that LLM-identified shocks are associated with smaller estimated multipliers than the narrative benchmark, with the difference linked in large part to differences in shock persistence and endogeneity.