Stock price prediction based on Recurrent Neural Networks

With the advent of the artificial intelligence era and the increasing demand for financial data analysis, Deep Learning (DL) has become an application frontier in the financial field. One of the application that has received widespread attention is the prediction of stock prices. The stock market is a complex nonlinear dynamic system, and predicting stock prices has consistently been a central area of research within the financial domain. The stock market is affected by many factors such as the economic situation, political environment, national policies, investor psychology, and other external markets. Its internal change rules are extremely complex, and conventional time series models face challenges in effectively addressing such intricate nonlinear problems. With the gradual development of Neural Network technology and the increase in the amount and availability of financial market data, DL has become the forefront of stock price prediction applications. Based on this, in this thesis, we explore and analyse the application of Recurrent Neural Networks (RNN) for the task of stock price prediction. The main contributions of this thesis can be divided into two parts. In the first part, we first propose an in-depth review of the most newsworthy DL-based models for stock price prediction recently proposed in the literature. Then we consider one of the most interesting of these works as a case study and we reproduce the results, uncovering some significant issues in the data pre-processing. We also extend the proposed results by testing with more advanced RNN architectures. In the second part, we proposed a novel architecture that relies on the distinguishing dynamic behaviour of the time series used as input for the models in the stock price prediction task. We dubbed this novel architecture multi-parallel-RNN. The obtained experimental results help us to analyse the strengths and weaknesses of the proposed approach, showing that the proposed architecture is able to outperform the considered baselines, improving the performance on the stock price prediction task.