Gilberto Boaretto

Working papers

Forecasting inflation using disaggregates and machine learning (with Marcelo C. Medeiros)

This paper examines the effectiveness of several forecasting methods for predicting inflation, focusing on aggregating disaggregated forecasts – also known in the literature as the bottom-up approach. Taking the Brazilian case as an application, we consider different disaggregation levels for inflation and employ a range of traditional time series techniques as well as linear and nonlinear machine learning (ML) models to deal with a larger number of predictors. For many forecast horizons, the aggregation of disaggregated forecasts performed as well as survey-based expectations and models that generate forecasts directly from the aggregate. Overall, ML methods outperformed traditional time series models in predictive accuracy, with outstanding performance in forecasting disaggregates. Our results reinforce the benefits of using models in a data-rich environment for inflation forecasting, including aggregating disaggregated forecasts generated from ML techniques, mainly during volatile periods. Starting from the COVID-19 pandemic, the random forest model based on both aggregate and disaggregated inflation achieved remarkable predictive performance at intermediate and longer horizons.

What news and social media tell us about future inflation? (with Marcelo C. Medeiros, Marcelo Fernandes, and Thiago Milagres)

We construct forward-looking indexes for inflation based on tweets and newspaper articles employing a supervised machine-learning approach. Using Brazilian data, we verify that the news-based indexes were able to anticipate long-term trends as well as capture short-term movements of the accumulated inflation over 3, 6, and 12 months ahead at various periods. Furthermore, the proposed indexes could improve inflation forecast performance. More specifically, for short horizons (3 and 6 months ahead), a bias correction model for the median of available survey-based expectations benefits from including news-based indexes. On the other hand, for longer-term inflation (12 months ahead), models incorporating a large number of predictors can also be improved by incorporating the indexes. Thus, considering indexes from social media and news sources can improve inflation forecasting. The intuition for the result is that it pays to consider a broader set of information than solely that resulting from survey-based expectations that account only for experts' opinions.

Work in progress

Time-varying bias-corrected average forecast (with Marcelo C. Medeiros)

We propose a bias correction for the average of a set of individual inflation expectations considering the possibility that intercept and slope biases may vary over time. We proceed in two ways. Firstly, we consider estimations based on rolling windows. Secondly, we employ a state-space model to obtain time-varying intercept and slope biases using the recursiveness of the Kalman filter. The latter approach has the advantage of circumventing the discretionary choice of the rolling window size. We also proceed with estimations based on extended windows, a procedure that is close to what has been done in the literature. We achieved good forecast performance for models based on short rolling windows for intermediate forecast horizons. In turn, a state-space model that includes corrections for intercept and slope biases varying over time tends to perform slightly worse than procedures based on rolling windows.

Should Central Banks React to Supply Shocks? Evidence from Brazil’s Core and Headline Inflation (with Antonio Bergallo and Raphael Ludwig)

We assess the drivers of Brazilian core inflation using a large structural Bayesian Vector Autoregression (BVAR) framework that assumes the target variables are influenced by a small number of common shocks. Our findings suggest that core inflation in Brazil is driven primarily by supply shocks over the period from 2006 to 2024. This result is somewhat unexpected, given that oil and food prices are excluded from the core inflation index, suggesting the presence of second-round and spillover effects in the economy. Moreover, we find no evidence that global supply chain disruptions significantly contribute to the post-pandemic inflation surge in Brazil. Our results support the view that central banks should respond to supply shocks.

Empirical asset pricing via factor-augmented models (with Caio Almeida, Jianqing Fan, and Marcelo C. Medeiros)

This research investigates the use of a model that combines both observed and unobserved factors and adds the effect of crossed idiosyncratic errors in stock return forecasting and portfolio management.

DSGE estimation via Hamiltonian Monte Carlo (with Daniel Coutinho)

DSGE models are frequently estimated by Markov Chain Monte Carlo (MCMC). Hamiltonian Monte Carlo (HMC) is a method that has not been used to estimate DSGE models. We explored how to implement that by employing Julia programming language.

Publications

DSGE Estimation using Generalized Empirical Likelihood and Generalized Minimum Contrast – with Márcio Laurini, Entropy 27(2), 2025.

We investigate the performance of estimators of the generalized empirical likelihood and minimum contrast families in the estimation of dynamic stochastic general equilibrium models, with particular attention to the robustness properties under misspecification. From a Monte Carlo experiment, we found that (i) the empirical likelihood estimator — as well as its version with smoothed moment conditions — and Bayesian inference obtained, in that order, the best performances, including misspecification cases; (ii) continuous updating empirical likelihood, minimum Hellinger distance, exponential tilting estimators, and their smoothed versions exhibit intermediate comparative performance; (iii) the performance of exponentially tilted empirical likelihood, exponential tilting Hellinger distance, and their smoothed versions was seriously compromised by atypical estimates; (iv) smoothed and non-smoothed estimators exhibit very similar performances; and (v) the generalized method of moments, especially in the over-identified case, and maximum likelihood estimators performed worse than their competitors.

Services inflation dynamics and persistence puzzle in Brazil: a time-varying parameter approach – with Cleomar Gomes, Applied Economics 51(6), 2018.

This article analyses services inflation dynamics in Brazil, focusing on the Services Inflation Persistence Puzzle, for monthly data from January 2004 to February 2016. We apply a time-varying parameter (TVP) approach, via a Kalman filter, to estimate hybrid Phillips curves and compare inflation inertia for tradable goods and services inflation. Aggregate Brazilian Extended Consumer Price Index inflation serves as a benchmark. To justify the TVP analysis, parameter instability and structural change tests are implemented, based on OLS and Generalized Method of Moments (GMM) frameworks. The main results are as follows: (i) the TVP approach is relevant due to observed instability in some parameters estimated; (ii) inflation expectation coefficients are higher than lagged inflation in all calculations, but inflation inertia is not negligible at all; (iii) services inflation persistence ranges from 27 to 36%, whereas tradable goods inflation persistence ranges from 36 to 47%, providing evidence of the Services Inflation Persistence Puzzle in Brazil; (iv) from 2009 onwards an increase in one percentage point in real wages raises monthly services inflation rate by 0.02 to 0.03 percentage point; (v) there is evidence that cost-push pressures, due to wage increases in the service sector, are more important to explain services inflation than demand pressures from early 2009 to mid-2014.

Inflation and Relative Price Variability in Brazil: a time-varying parameter approach – with Cleomar Gomes, Economics Bulletin 38(4), 2018.

This paper aims at analyzing the relationship between inflation and relative price variability (RPV) in Brazil, for the period spanning from July 1999 to May 2017. State Space models, together with Kalman Filter, are used as main econometric methodology. As a benchmark exercise, ARDL-OLS and GMM estimations are applied, enabling to compare time-fixed and time-varying parameter approaches, and to examine the impact of potential endogeneity on the results. A GARCH-type model is also assessed to account for inflation variability. Results show that: i) the correlation between inflation and RPV is positive, but with statistical significance only in some months up to 2003 and from 2015 on; ii) the pass-through from inflation to RPV seems to increase when inflation increases, especially in the beginning of 2015; iii) the time-varying parameter approach proves to be relevant due to instabilities found in time-fixed estimations; iv) inflation variability is statistically significant and affects relative price variability throughout the whole period analyzed.

Google Sites

Report abuse