Deep learning on SPY stock prices using improved multi-head LSTM

Abstract

Stock price prediction has been a widely pursued topic by researchers in recent years due to the great impact that significant research can have on the economy. LSTM is commonly used for stock price prediction as it has strong time series predictive capabilities. However, it is limited by its loss function, which only takes one parameter (predicted stock price) into account. This paper proposes a multivariate multi-step, vector output predictive model using LSTM, with both the stock price and the relative return as inputs of the neural network. Furthermore, a novel loss function that combines both the standard mean squared error, and the relative return mean squared error to hone accuracy is introduced. The model’s predictive capabilities are demonstrated on the S&P 500 Index. This improved LSTM model reaches a test MSE of 0.0076, which is a result that is significantly stronger than results demonstrated by standard LSTM stock prediction networks, and which outperforms most of the LSTM models mentioned in the literature.