Which One Predicts Better?: Comparing Different GDP Nowcasting Methods Using Brazilian Data
Peer Reviewed
Date issued
Jul 2023
Subject
Inflation;
Potential Output;
Output Gap;
Macroeconomic Policy;
Development Bank;
Debtor Finance;
Prediction Market;
Gross Domestic Product;
Economy;
Machine Learning;
Learning Strategy;
Industrial Productivity;
Learning;
Economic Development
JEL code
C53 - Forecasting and Prediction Methods • Simulation Methods;
C45 - Neural Networks and Related Topics;
E17 - Forecasting and Simulation: Models and Applications
Country
Brazil;
Brazil
Category
Discussion Papers
The objective of this paper is to develop a basic framework for the implementation of a GDP nowcasting strategy using Brazilian data. Our goal is to identify a scalable strategy that allows us to project the Brazilian GDP in real time at any point during the current quarter. In the paper we detail the survey of classical techniques and also of techniques usually known by market practitioners as "machine learning methods". We survey the literature since the first work on estimating business cycles and document the evolution of this literature until the insertion of machine learning methods. Additionally, we perform backtesting exercises, estimate several candidate models for GDP nowcasting. Finally, we evaluate the forecasting power of all models against a naive model and a market expectations model. We demonstrate that a combination of machine learning models based on the distance of forecasts to the average market expectations defeats the fully informed market expectations, while the same is not possible for selected classical nowcasting models.
