Identifying Monetary Policy Shocks in Newspapers Using GPT

Why It Matters

The approach offers a scalable, less noisy proxy for ECB policy shocks, improving macroeconomic analysis and policy monitoring, especially in periods of financial turbulence.

Key Takeaways

• GPT classifies newspaper articles with up to 95% relevance accuracy.
• Five‑point surprise scale yields 89.8% agreement with human coders.
• Newspaper‑based shock series correlates 0.6‑0.8 with high‑frequency shocks.
• Only 14% of identified surprises stem from information effects.
• Method stays robust during financial stress, outperforming traditional HFI.

Pulse Analysis

Identifying monetary‑policy shocks has long relied on structural models or high‑frequency market data, both of which grapple with endogeneity and information‑effect biases. Narrative identification—tracing policy surprises in textual sources—offers a complementary lens, but traditional human coding is labor‑intensive and prone to inconsistency. Recent advances in generative AI now enable automated, reliable text analysis, opening the door to large‑scale, real‑time shock measurement.

The study by Betz et al. (2026) combines these strands by harvesting articles from eleven leading European newspapers immediately after each ECB meeting and feeding them to a suite of OpenAI models. After a deterministic pre‑filter, the LLM tags relevance (91‑95% accuracy) and assigns a surprise rating on a –2 to +2 scale, matching human judgments 89.8% of the time. When aggregated, the newspaper‑based surprise series aligns closely (0.6‑0.8 rolling correlation) with the benchmark high‑frequency shock series, yet it diverges during the Global Crisis in ways that better reflect the narrative context. Crucially, only about 14% of the identified surprises are driven by information effects, a stark contrast to the roughly 50% reported for conventional high‑frequency measures.

For researchers and policymakers, this hybrid LLM‑driven narrative tool provides a more robust, cost‑effective way to monitor ECB policy dynamics, particularly when market data are noisy or when financial stress alters traditional signal pathways. It also demonstrates that large language models can reliably replicate nuanced human judgment in economic research, suggesting broader applications for AI‑assisted textual analysis across macro‑financial domains. Future work may expand the newspaper pool, refine prompting strategies, and integrate cross‑currency comparisons to further enhance the precision and universality of AI‑generated shock metrics.