Enhancing Stock Return Predictions in the S&P 500: A Comparative Study of Random Forests and LSTM Networks
Leveraging Multi-Feature Approaches for Superior Intraday Trading Performance from 1990 to 2018
In this article, we tested two methods, random forests and LSTM networks (CuDNNLSTM), to see how well they can predict future stock movements in the S&P 500 index. We looked at data from January 1993 to December 2018 for intraday trading. Our approach included considering closing prices, opening prices, and intraday returns for multiple stocks.
For our trading strategy, we followed the methods described by Krauss and colleagues in their 2017 and 2018 studies. Each trading day, we chose to buy the 10 stocks that were likely to perform well and sell short the 10 stocks that were likely to perform poorly, based on their intraday returns. We invested the same amount of money in each stock.
Our results showed that using the multiple stock features, we achieved a daily return of 0.64% with LSTM networks and 0.54% with random forests, before considering transaction costs. These results were better than the single-feature approach used in the studies by Fischer & Krauss (2018) and Krauss et al. (2017), where they only considered daily returns based on closing prices. In those studies, the daily returns were 0.41% and 0.39% for LSTM and random forests, respectively.
