Financial time series forecasting using artificial neural networks

• Autor(a) principal: Cardoso, Luís Gil Miguéns
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10071/21560
• topic: Time series forecasting; Artificial neural networks; Stock market; Long short-term memory; Stacked autoencoder

This study builds an artificial neural network framework with the use of stacked autoencoders (SAE) to extract deep denoised features, and long short-term memory (LSTM) to generate forecasts for the next-day adjusted closing price of S&P500. Data for seven different stock indices, technical indicators, and macroeconomic variables is used to train three different models: a 'price model' which predicts the next-day price, a 'change model' which predicts the relative change in price, and a ’binary model’ which predicts the probability of a price increase. The models were judged based on predictive accuracy and profitability. Results show the models either fail to generalize well or fall prey to a vicious minimum approximating a naive predictor. Furthermore, the models appear particularly poor at predicting breaks in the series, likely due to their infrequency. This might provide evidence supporting the efficient market hypothesis.