Stock Price Analysis with Deep-Learning Models

Novel artificial intelligence prediction algorithms use deep learning techniques, i.e., recurrent neural networks and convolutional neural networks, to predict financial time series. Also, autoencoders have gained notoriety to extract features from latent space data and decode them for predictions. This paper compares several deep learning architectures with different combinations of long short-term memory networks and convolutional neural networks. Autoencoders are implemented within these networks to find the best model performance for financial forecasting tasks. Four different architectures were trained with stock market data of four companies (AMD, ResMed, Nvidia, and Macy's) from 2010 to 2020. Without autoencoder, the long short-term memory network architecture achieved the best performance for all companies, obtaining a mean squared error of 0.004 for AMD stocks by applying 10-fold nested cross-validation. The results show that long short-term memory networks are very well suited for prediction tasks using a simple deep-learning architecture.