Application of LSTM, GRU and ICA for stock price prediction

Abstract

This paper attempts to provide an optimal model for the prediction of stock prices for t + 5th day and consequently provide a daily buying/selling strategy for the Standard’s and Poor’s 500 Index. A performance comparison between LSTM, GRU, ANN and SVM model has been made and an optimal model has been outlined. Training and prediction data spanned over 12 years from 2000 to 2017. Fifty technical indicator-based attributes were calculated and appended to the open, high, low, close and volume (OHLCV) data for S&P500, each attribute value was converted into a relative standard score followed by minimax scaling and dimensionality reduction, through ICA. Performances of different models on this dataset were then compared using self-defined metrics like optimism and pessimism ratio and returns ratio. The LSTM model proved to outperform the other models with a return of 400% greater than the hold and wait strategy and R2 score of 0.9486.