The impact of price disaggregation on the estimation of aggregate trend inflation and volatility in Brazil
Cleomar Gomes da Silva, Júlio Fernando Costa Santos, Sinara do VallePurpose
This article aims to empirically analyze the behavior of trend inflation and volatility in Brazil.
Design/methodology/approach
We apply the Uni-Multivariate Unobserved Components/Stochastic Volatility Outlier-Adjustment (MUCSVO) methodology. The models are estimated at the monthly frequency from August 1999 to December 2024, combining sectoral disaggregation with a one-sided latent-seasonal filtering procedure that avoids forward-looking contamination. The resulting measures are evaluated in-sample and through a recursive pseudo-out-of-sample forecasting exercise at the quarterly frequency. The estimations are based on disaggregated sectoral data corresponding to nine components of the Consumer Price Index (IPCA) basket, complemented by additional analyses for monitored and free market prices, as well as the core inflation measure.
Findings
The results show that part of the increase in transitory volatility estimated by univariate models after 2010 is reclassified as idiosyncratic sectoral noise rather than an aggregate phenomenon. Multivariate models also assign high posterior outlier probabilities to major stress episodes, notably the 2002 presidential election crisis and the COVID-19 pandemic. In-sample, the nine-component model reduces posterior uncertainty by about 30% relative to the univariate benchmark, a gain interpreted as model-conditional posterior precision rather than external validation of the latent trend. Out-of-sample, the strongest forecasting gains come from a parsimonious two-component specification based on monitored and free-market prices.
Originality/value
Relative to Stock and Watson (2016), our contribution is not merely to apply the multivariate framework to an emerging economy, but to adapt it to a setting marked by stronger sectoral heterogeneity, regulated-price adjustments, and recurrent extreme shocks. We estimate the models at the monthly frequency from August 1999 to December 2024, combine sectoral disaggregation with a one-sided latent-seasonal filtering procedure that avoids forward-looking contamination, and evaluate the resulting measures both in-sample and through a recursive pseudo-out-of-sample forecasting exercise at the quarterly frequency. The empirical analysis considers three levels of disaggregation: the nine official IPCA groups, the monitored-versus-free-price decomposition, and the Core EX1 structure.