Ensaios em macroeconomia: previsão macroeconômica, risco bancário e preferências do Banco Central

• Ano de defesa: 2022
• Autor(a) principal: Jesus, Diego Pitta de
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal da Paraíba; Brasil; Economia; Programa de Pós-Graduação em Economia; UFPB
• Palavras-chave em Português: Previsão macroeconômica; Machine learning; Índice de sentimento; Ata do copom; Relatório da inflação; Insolvência bancária; Cluster; Sentimento bancário; Análise de sentimentos; Função de reação; Sentimento da política fiscal; DSGE; Macroeconomic forecast; Sentiment index; Copom minutes; Inflation report; Bank insolvency; Banking sentiment; Sentiment analysis; Reaction function; Sentiment of fiscal policy; CNPQ::CIENCIAS SOCIAIS APLICADAS::ECONOMIA
• Link de acesso: https://repositorio.ufpb.br/jspui/handle/123456789/24759

The aim of the paper is to verify if the tone of the reports produced by the Central Bank of Brazil contain information that can be used to improve the precision of the projections of the macroeconomic indicators for a quarter ahead. Thus, we built predictors for inflation and GDP growth obtained from textual analyzes of the Copom Minutes and Inflation Report. For the creation of sentiment scores, we used a traditional fixed-lexicon dictionary approach and a new approach that uses machine learning to generate a time-varying dictionary. Next, we test the predictive power of the new variables for macroeconomic indicators for a period ahead. We also tested whether these new predictors are able to improve the performance of predictive models. The results show that the best predictions were obtained with the models that used the time-varying dictionary textual score series. The fact happened because this type of dictionary is capable of incorporating new terms that appear in the reports. We also found that market forecasts of average GDP growth can be improved with sentiment scores. But this was not verified for inflation. The main motivation of this paper is to use machine learning techniques to build a new insolvency risk rating metric for banks traded on B3. Then, a set of prediction models will be used to project the risk classification of these institutions. Conventionally, the literature analyzes the risk of bank insolvency based on accounting data and macroeconomic variables. In addition to these variables, this work will build a series of sentiment of the bank’s manager, via quarterly reports (ITR), and this will be used to improve the accuracy of bank risk forecasts. The results indicate that the banking risk classification, by the k-means algorithm, was able to classify 17% of the sample in the highest risk group (1), while 83% of the sample was in the lowest bankruptcy risk group (0). Using the Z-score metric, we found that 65% of the sample is part of the low-risk group and 35% of the sample is part of the high-risk group. Thus, the k-means algorithm is more rigorous in classifying a bank in the highest risk category. Next, we use the data already described to project the risk of bank insolvency. The results of this step showed that the decision tree model presented the best performance for the sample of test. In addition, it was found that the inclusion of the banking sentiment variable was able to improve the performance of forecast models, especially when banking sentiment is constructed from a time-varying dictionary. The objective of the paper is to investigate whether the Central Bank has been reacting to fiscal policy sentiment. The variable that measures the polarity of fiscal policy was constructed through natural language processing and sentiment analysis of monthly public debt reports issued by the National Treasury. The sentiment index was inserted as a dependent variable in two approaches to achieve the objective of the work. The first is the estimation of a traditional central bank reaction function. The second is the estimation of a DSGE model to estimate reaction functions and thus produce inferences about the effect of fiscal policy sentiment on monetary policy behavior. The main results suggest that fiscal policy sentiment has explicitly entered the monetary policy decision-making process in Brazil, indicating a possible scenario of fiscal dominance.