PPGI | (92) 3305-1181 Ramal 1193 / (92) 99128-5875 |
Convite 341ª Defesa de Mestrado - Janderson Borges do Nascimento
Sexta-feira, 13 Setembro 2019,  2:00 -  5:00
Contato: Secretaria do PPGI (Este endereço de email está sendo protegido de spambots. Você precisa do JavaScript ativado para vê-lo.)

A Coordenação do Programa de Pós-graduação em Informática PPGI/UFAM tem o prazer de convidar toda a comunidade para a sessão pública de apresentação da 341ª Defesa de Dissertação de Mestrado: 

TÍTULO: StockNet: A Multivariate Deep Neural Architecture for Stock Prices Prediction

CANDIDATO: Janderson Borges do Nascimento


Prof. Dr. Marco Antônio Pinheiro de Cristo - PPGI/UFAM (Presidente) 
Prof. Dr. Rafael Giusti - PPGI/UFAM (Membro Interno) 
Prof. Dr. Moisés Gomes de Carvalho - IComp/Ufam (Membro Externo) 


Stock price forecasting is an inherently difficult problem. According to the efficient market hypothesis financial prices are unpredictable. However, a great number of machine learning methods have obtained consistent results on anticipating market movements. Most recent time-series prediction methods tried to to predict the polarity of prices, that is, whether they will have risen or fallen in relation to the previous temporal step. That such approaches seems not feasible in real scenarios, since it predicts only the price polarity, preventing a consistent planning of gains and losses while it avoids inferring how much each operation will yield. Most of these methods use only Recurrent Neural Networks, but recent advances in temporal convolutional networks also may prove to be promising in prediction of general time-series, making possible better and faster predictions with easier to train and low memory usage models. Recent hybrid architectures have also obtained important results using additional unstructured information from finance news. We propose a novel deep neural architecture to predict stock prices based on Temporal Convolutional Networks and built upon on a state of the art acoustic model for voice synthesis. Experimental results show that our model can consistently improve individual stocks prediction when compared to traditional methods.


Local Auditório 01 - ICompTec

Produzido por :