A More Robust Monetary Policy Shock: A Natural Language Processing Approach of Thailand's Monetary Policy Reports

This study develops a more robust strategy for identifying monetary policy shocks in Thailand by incorporating both quantitative forecasts and qualitative sentiment extracted from central bank communications, specifically Monetary Policy Report. Building on Romer and Romer, 2004 framework, I apply Ridge regression to estimate the policy rate using a rich information set, combining macroeconomic forecasts with sentiment indices derived through topic modeling and dictionary-based analysis of Monetary Policy Reports from 2013 to 2023. The residuals from this model are interpreted as exogenous monetary policy shocks. Impulse response functions (IRFs), estimated via a Vector Autoregression (VAR) model, show that a contractionary shock leads to a decline in inflation and appreciation of the real effective exchange rate, consistent with theoretical expectations. However, real GDP growth initially rises and unemployment briefly falls, suggesting short-run deviations possibly due to anticipatory effects or delayed transmission. These dynamics gradually align with standard macroeconomic theory over time. Compared to shocks identified using ordinary least squares (OLS), Ridge-based shocks generate more stable and plausible responses across macroeconomic variables. Exogeneity tests confirm that the shock series is not predicted by lagged macroeconomic indicators. These findings highlight the value of integrating natural language processing and machine learning to enhance monetary policy analysis in emerging market economies.