More Is More? Bayesian VARs for forecasting and monetary policy in Brazil

This thesis applies large-scale Bayesian Vector Autoregressive (BVAR) and FactorAugmented VAR (FAVAR) models to the Brazilian economy, drawing on 35 monthly time series from March 1999 through September 2025. The first chapter compares the out-of-sample forecasting accuracy of three BVAR specifications (Small, n = 4; Medium, n = 16; Large, n = 35) against univariate benchmarks, frequentist VARs, a random walk, and the Central Bank of Brazil’s Focus survey of professional forecasters. The hierarchical prior framework of Giannone, Lenza, and Primiceri (2015), combined with the Pandemic Priors of Cascaldi-Garcia (2022), estimates the shrinkage hyperparameters for each model dimension. The Medium BVAR forecasts best among the statistical models, beating the random walk for inflation and output; the Large BVAR, with thirty-five variables, adds nothing at this horizon. Judged on the joint predictive density, the Medium and Small BVARs outperform every statistical benchmark, and the Model Confidence Set rejects the largest system once the field narrows to the three BVARs. Only the univariate AR beats Focus, and only for output; at twelve months the Model Confidence Set keeps the survey-augmented Medium BVAR and the professional consensus in the same 95% superior set. The second chapter asks whether a broader information set, in the BVAR and in a FAVAR estimated on the same dataset, attenuates the well-known price puzzle of recursive identifications: the counter-intuitive rise of prices after a contractionary monetary policy shock. As the information set expands, the puzzle attenuates: the sixteen-variable Medium removes most of it, and the FAVAR goes furthest, pushing the price response below zero, consistent with an omitted-information interpretation rather than a structural feature of the Brazilian economy.