Extended Abstract

This paper examines how sentiment expressed in central bank annual reports influences stock market behavior through both direct and indirect channels. Using FinBERT, a financial domain-specific language model, we analyze almost three decades of annual reports from 13 central banks and construct a stock market related sentiment indicator. We classify sentence-level references to stock markets and inflation as positive, neutral, or negative, constructing Central Bank Sentiment (CBS) indicators for each topic. The empirical analysis uses an enhanced Fama–French model estimated via pooled OLS, fixed and random effects, as well as Prais–Winsten regressions with panel-corrected standard errors. The results consistently show that central bank sentiment exerts a significant and robust influence on stock returns. Positive sentiment regarding the stock market is generally associated with higher returns, while negative sentiment amplifies sensitivity to market risk. Robustness checks include: i) leave-one-out analysis to ensure results are not driven by a single central bank, ii) subsample analysis (1995–2007 vs. 2008–2022) to capture post-GFC structural changes and lastly, iii) quantile regressions (10th–90th percentile) to assess heterogeneous effects across market conditions. Across all robustness checks, we find that negative sentiment exerts a stronger impact than positive sentiment on stock returns and market risk. These findings extend the 'central bank information channel' literature and validate the robustness of large language models in analyzing central bank communications across economic cycles and regions.