Estudo em séries temporais financeiras utilizando redes neurais recorrentes

• Autor(a) principal: Matsumoto, Daniel Kazuyuki Fugii
• Data de Publicação: 2019
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da Universidade Federal de Alagoas (UFAL)
• Texto Completo: http://www.repositorio.ufal.br/handle/riufal/6813
• topic: Análise de séries temporais; Séries financeiras; Domínio de tempo; Redes neurais (Computação); Redes neurais recorrentes; Mercado financeiro; Autoregressive integrated moving average; Long short-term memory; Time series analysis; Financial series; Time domain; Neural networks (Computing); Recurrent neural networks; Financial market; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO

Understanding the behavior of financial time series is very importance to forecast future values and make efficient decisions. This area is often the target of studies and new proposals in the attempt of its modeling, since it is very volatile in response to the dynamism of the market. This volatility is defined by the high number of variables and different sources of information, a low signal-to-noise ratio, and its predictions fail to generalize the problem. In addition to existing relations between unknown and random exogenous variables, capable of strong influence on the current structure of the data. However, just as incorrect information can cause great misfortune, more accurate information can generate proportional profits, thus ensuring that this problem is the subject of several studies throughout history. For this purpose a predictive model was designed using Recurrent Neural Networks (LSTM), fed from its historical dataset (time series) to forecast the oppening value trend of variation for an asset (stock). Thus a variety of experiments were performed and their results analyzed against the widely used metrics, like autoregressive models (ARIMA) to define a basis for comparisons and validate the model capacity. The results obtained were satisfactory, obtaining an acuracy of up to 74:50% by predicting the sign of the next opening value, inferring whether or not a given stock will rise in the future.