Application of Deep Learning Methods in Predicting Stock Market Direction

Abstract

. This paper selects daily data from the S&P 500 from 2013 to 2024 and constructs a binary prediction box for stock market direction using average market return as the core indicator. Logistic regression is used as the baseline method to characterize linear relationships; Long Short-Term Memory (LSTM) is used to model the dynamic dependency structure in time series. Empirical results show that LSTM has a higher prediction accuracy than logistic regression, but the AUC values of both models are close to stochastic levels, indicating that it is difficult to judge market direction based solely on historical price information. It also shows that although deep learning models have advantages in representing nonlinear features, the improvement in predictive ability in noisy financial environments remains limited. This study provides empirical evidence for the selection of methods for predicting overall market trends.