DEEP LEARNING APLICADO À PREDIÇÃO DE TENDÊNCIAS NO MERCADO DE AÇÕES

Abstract

Aiming to achieve the goal of training neural networks to enable them to identify trends in the price of a stock, data were collected on the stocks of 7,145 companies during the period of one month. These data were transformed into candlestick charts that, using technical analysis concepts, were classified as being charts with high, low and consolidation tendencies and formed the bases used to train, validate and test ResNet and Xception network models. It was possible to conclude that both obtained satisfactory results, but with advantage to ResNet, that obtained greater precisions from the training phase until the final tests. In the final tests ResNet reached 87% of accuracy and F1 Score, while Xception reached 86%.