Using Natural Language Processing to Identify Monetary Policy Shocks

Identifying the causal effects of monetary policy is challenging due to the endogeneity of policy decisions. In recent years, high-frequency monetary policy surprises have become a popular identification strategy. To serve as a valid instrument, monetary policy surprises must be correlated with the true policy shock (relevant) while remaining uncorrelated with other shocks (exogenous). However, market-based monetary policy surprises around Federal Open Market Committee (FOMC) announcements often suffer from weak relevance and endogeneity concerns. This paper explores whether text analysis methods applied to central bank communication can help mitigate these concerns. We adopt two complementary approaches. First, to improve instrument relevance, we extend the dataset of monetary policy surprises from FOMC announcements to policy-relevant speeches by the Federal Reserve Board chair and vice chair. Second, using natural language processing techniques, we predict changes in market expectations from central bank communication, isolating the component of monetary policy surprises driven solely by communication. The resulting language-driven monetary policy surprises exhibit stronger instrument relevance, mitigate endogeneity concerns and produce impulse responses that align with standard macroeconomic theory.