Previsão da taxa de câmbio Real/Dólar norte-americano usando modelos de Deep Learning

Abstract

Exchange rate forecasting is one of the greatest challenges in economic and financial literature, holding strategic importance for policy formulation and decision-making aimed at efficient resource allocation. This dissertation investigates the forecasting of the Ptax exchange rate for the Brazilian Real against the US Dollar (BRL/USD) using Deep Learning (DL) models, evaluated over time horizons of 1, 10, and 20 business days ahead, and compared to the Random Walk (RW) model, which is widely recognized in the literature as a robust benchmark for exchange rate forecasting. Daily data series covering the period 2020 to 2024 were used, sequentially divided into training, validation, and test sets to simulate a practical application. The study adopted two approaches: a univariate approach, based exclusively on exchange rate data, and a multivariate approach, which incorporated economic and financial variables. For each forecast horizon and approach, DL architectures with one (1CR) or two recurrent layers (2CR) were tested. DL model selection was guided by error metrics on the validation set, followed by performance evaluation on the test set. The results show that, in the univariate approach, the predictions of the selected DL models did not outperform those of RW on the test set, a period in which the data is new to the models. Only on the 10-business-days horizon did the Modified Diebold-Mariano (DMM) test indicate statistical equivalence. However, for the 20- business-days horizon, a sharp deterioration in the performance of the selected DL models was observed, with error metrics at least 50% higher than those of the RW. On the other hand, in the multivariate approach, the DMM test, at the 10% significance level, indicated that the forecasts of the selected DL models performed statistically better than those of RWover the 20-businessdays horizon, in both architectures evaluated (1CR and 2CR). Consistently, the results of the Giacomini and Rossi (2010) fluctuation test and the Rossi and Sekhposyan (2016) rationality test indicated, respectively, in the time windows, equal or superior accuracy of the selected multivariate DL models’ forecasts, and lower bias and irrationality in errors compared to the RW. It is concluded that, although the DL models did not outperform the benchmark in the very short term, the inclusion of daily economic and financial variables provided significant gains in predictive performance over longer horizons, even in the short term. The results reinforce the potential of multivariate DL models to improve exchange rate forecasting, contributing both to the academic literature and to practical applications in the financial market.