Seeing the forest for the trees: Using hLDA models to evaluate communication in Banco Central do Brasil

BIS Working Papers  |  No 1021  | 
14 June 2022

This paper was produced as part of the Final Conference of the BIS-CCA Research Network on "Monetary policy frameworks and communication (2019-2022)".



There is a widespread consensus that central bank communication influences the expectations of economic agents, and that increasing transparency enhances the effectiveness of monetary policy. Central bank communication has moved beyond simply informing agents on the perceived current and future state of the economy to becoming a key instrument to anchor inflation expectations. Therefore, it is essential to assess whether the central bank is efficiently communicating the intended signals to market and economic agents.


We use computational linguistic techniques to analyze the content and tone of the statements and minutes of the Monetary Policy Committee (Copom) of the Banco Central do Brasil. We construct sentiment indexes that measure the perception of the Copom on inflation, economic activity and uncertainty, based on dictionary methods applied to a hierarchical Latent Dirichlet Allocation (hLDA) model combined with feature selection techniques. This assures that every topic in the tree contains meaningful words for proper analysis, favors interpretability, and provides relations between themes without intervention by the researcher. The resulting sentiment indexes are compared with actual data to assess whether the Copom documents reflect the state of the economy. The model also allows for evaluating the coherence between statements and minutes.  


The Copom statements and minutes properly communicate the actual state of the economy, although the correlation between sentiment indexes and economic variables was affected by significant changes in Banco Central do Brasil's communication in July 2016. This event affected the correlation between the indexes and observables and their volatility, which can be partially attributed to changes in the average number of words dedicated to topics associated with a specific subject. Even though statements do not share the same information with all details present in minutes, they both deliver the same message.


Central bank communication is a key tool in managing inflation expectations. This paper proposes a hierarchical Latent Dirichlet Allocation (hLDA) model combined with feature selection techniques to allow an endogenous selection of topic structures associated with documents published by Banco Central do Brasil's Monetary Policy Committee (Copom). These computational linguistic techniques allow building measures of the content and tone of Copom's minutes and statements. The effects of the tone are measured in different dimensions such as inflation, inflation expectations, economic activity, and economic uncertainty. Beyond the impact on the economy, the hLDA model is used to evaluate the coherence between the statements and the minutes of Copom's meetings.

JEL classification: E02, E21, E22.

Keywords: communication, monetary policy, latent dirichlet allocation, Brazil, Central Bank.