Redes neurais ELU feed-forward aplicadas a previsão de preços de ativos em bolsas de valores

Abstract

One of the most fashionable forms of financial investment available in nowadays economy is the buying and selling of stocks. Currently the vast majority of stock transactions are already done automatically by specialist software, but there is still a lot of room to develop specific decision-making methods for that area. The present work tests an approach to predict the rise and fall of stock prices using neural networks that although not new, has had its viability increased in the last years thanks to the advances in the algorithms of training and parallelization using graphics cards. In this work, we compare feed-forward neural network architectures with different hyperparameter configurations, specifically using ELU-type neurons, applied to the daily minimum price forecast of the S&P 500 index. For the programming of the algorithms, the Python language was used, along with the framework for Deep Learning Tensorflow, and some libraries that simplify the process of data manipulation (Pandas) and network structuring (Keras). After analyzing and choosing the best architecture, an optimization process is shown in which it is verified that the performance of the network is improved with the addition of new layers of neurons.