StockNet: A Multivariate Deep Neural Architecture for stock prices prediction

• Ano de defesa: 2019
• Autor(a) principal: Nascimento, Janderson Borges do
• Outros Autores: http://lattes.cnpq.br/5357554382847406
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal do Amazonas; Instituto de Computação; Brasil; UFAM; Programa de Pós-graduação em Informática
• Palavras-chave em Português: Redes neurais (Computação); Processamento de linguagem natural (Computação); CIÊNCIAS EXATAS E DA TERRA: CIÊNCIA DA COMPUTAÇÃO: TEORIA DA COMPUTAÇÃO: COMPUTABILIDADE E MODELOS DE COMPUTAÇÃO; Stocks Forecast; Neural Networks; Natural Language Processing; Time-Series Prediction; Redes Neurais; Processamento de Linguagem Natural; Previsão de Séries Temporais; Previsão de Ações
• Link de acesso: https://tede.ufam.edu.br/handle/tede/7409

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 attempt to predict prices polarity, that is, whether prices have increased or fallen compared to the last time-step. Such approaches are inefficient in real scenarios, as forecasting price polarity alone makes financial planning a hard task, due to the fees and operation costs. 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 better predictions with easier to train models. Recent hybrid architectures have also obtained important results using additional unstructured information from financial 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.