Stock Price Prediction by Using Hybrid Sequential Generative Adversarial Networks

Abstract

A significant application of machine learning in the financial field is stock price prediction. Investors can obtain a useful investment reference from the result of a stock prediction model, and for the whole financial system, it can optimize resource allocation. Stock future trend prediction is mainly divided into fundamental and technical analyses. Before the boom of machine learning, the ARIMA model was most used in stock price prediction. In recent years, according to the development of machine learning, the state-of-art algorithms of machine learning such as Long Short-Term Memory (LSTM) and Generative Adversarial Networks (GANs) were used to predict stock price. In previous research, however, only one single model had been used for this task. In this work, we used the hybrid sequential GANs model, it set different Recurrent Neural Networks(RNN, LSTM, GRU) in the two components(Generator and Discriminator) of GANs. We designed three training strategies to train our model by using the data of the S&P 500. There are two evaluation methods in this work: different loss functions(RMSE, MAE) and the accuracy of classification on buy, hold, sell strategy. It is proved through experiments that hybrid sequential GANs has a better performance in the stock prediction than the previous single algorithm prediction research.