Testing the Ability of Deep Learning Models LSTM and Transformers to Predict the Price Behavior of the S&P 500 Index

Abstract

This study aims to test the ability of deep learning models (LSTM and Transformers) to predict the price behavior of the S&P 500 index, with a focus on comparing the performance of both models under different market scenarios. The study adopted a quantitative experimental approach, using daily data of the S&P 500 index from January 2015 to December 2023, with a total of 2,264 trading days. Two main models were constructed: a three-layer LSTM model and a Transformer model customized for financial data. The performance of both models was evaluated using four main metrics: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Directional Accuracy. Paired samples t-test and Wilcoxon test were used to verify the statistical significance of differences. The results showed a clear superiority of the LSTM model over the Transformer in all performance metrics. The LSTM achieved an RMSE of 79.86 compared to 145.19 for the Transformer, and recorded a directional accuracy of 51.22% compared to 48.17% for the Transformer. Statistical tests showed significant differences between the performance of the two models at the 0.05 significance level