Aprimorando previsões de retornos em ativos financeiros: Uma abordagem híbrida com arima e o modelo timesnet

Abstract

This Undergraduate Thesis aims to analyze the performance of the classical econometric model ARIMA (Box and Jenkins, 1976) and the deep learning model TimesNet (Wu et al., 2022), applied to the forecasting of financial time series, with attention to the relationship between signal and noise, a central element that defines the degree of predictability and the linear or nonlinear nature of the data. The research evaluated the effectiveness of each approach in terms of statistical accuracy. The methodology included the use of widely recognized metrics, such as RMSE, MAE, and MAPE, applied to assets representing mature and emerging economies, including stock market indices (United States, Brazil, Germany, the United Kingdom, and Japan) and the gold commodity. Empirical results showed that the ARIMA model, based on a linear and parsimonious structure, performed better in short-term horizons and in series with low signal and high noise, confirming its efficiency in environments with lower predictability. On the other hand, TimesNet, by employing deep learning mechanisms capable of capturing nonlinear patterns and complex dependencies, presented advantages only when combined with ARIMA, reinforcing the potential of hybrid models. Thus, it is concluded that the integration between classical models and advanced artificial intelligence techniques allows for the complementary exploration of linear and nonlinear signals, enhancing return forecasting and contributing to more assertive decision-making in the dynamic context of investment portfolio management.