Essays on inflation forecasting and monetary policy in Brazil

Resumo

This thesis consists of two independent chapters. The first chapter provides a comprehensive comparison between traditional econometric methods and modern machine learning (ML) algorithms for forecasting Brazilian inflation. We assess predictive performance across a wide range of horizons and under different macroeconomic regimes, explicitly distinguishing between periods of positive and negative output gap. The results demonstrate that tree-based ML models, particularly Random Forest, are robust performers for inflation forecasting, consistently achieving the lowest forecast errors across all horizons and throughout the entire sample, regardless of the prevailing economic regime. Boosted Regression Trees also perform especially well during periods of negative output gap. Both models outperform traditional econometric benchmarks and the Focus survey in the medium and long term, highlighting the advantages of flexible, nonlinear approaches for capturing the complexities of inflation dynamics. Furthermore, we find no significant change in variable selection as the forecast horizon extends, suggesting a more stable set of relevant predictors for Brazilian inflation than previously reported. Altogether, these findings reinforce and extend the existing literature by demonstrating the robustness of ML methods, especially Random Forest, across different phases of the business cycle. The second chapter estimates a time-varying parameter vector autoregressive model with stochastic volatility (TVP-VAR-SV), following the methodology of Primiceri (2005), to analyze the evolution of Brazilian monetary policy since the mid-2000s. Employing the corrected estimation algorithm of Del Negro and Primiceri (2015), the study jointly models the dynamics of monetary policy responses to macroeconomic shocks (including output growth and the difference between twelve-month-ahead inflation expectations and the inflation target), as well as the response of these variables to shocks in the Selic interest rate, alongside the variances of these shocks. This approach enables a detailed decomposition between systematic components (based on a Taylor-type rule) and discretionary components of monetary policy. The empirical results show that the Central Bank of Brazil has broadly adhered to the Taylor principle throughout the entire sample period, including during Tombini’s administration. The impulse response functions (IRFs) indicate that both the transmission mechanism of monetary policy and the immediate response to shocks in output and inflation expectations have remained stable over time, with no significant changes detected. As a result, differences observed between policy regimes are mainly explained by discretionary actions rather than systematic rule changes—a conclusion further reinforced by counterfactual simulations.

Ficha do documento

Tipo

Tese

Ano

2025

Instituição

Fundação Getulio Vargas

Fonte

Repositório da FGV

Idioma

Inglês

Acesso

Acesso aberto

Identificador

oai:repositorio.fgv.br:10438/37295

Temas

Economia

Palavras-chave

Previsão da inflaçãoModelos de árvorePolítica monetáriaRegra de TaylorTVP-VARChoques estruturaisAnálise contrafactualMachine learningRandom forestInflation forecastingTree-based modelsMonetary policyTaylor RuleStructural shocksCounterfactual analysisEconomiaInflação - PrevisãoAprendizado do computadorÁrvores de decisãoModelos econométricos