Predicting daily stock price directions with deep learning models

Triparna Kundu, Eugene Pinsky,
Predicting daily stock price directions with deep learning models,
Machine Learning with Applications,
Volume 22,
2025,
100744,
ISSN 2666-8270,
https://doi.org/10.1016/j.mlwa.2025.100744.
(https://www.sciencedirect.com/science/article/pii/S2666827025001276)
Abstract: In this paper, deep learning models like LSTM, CNN and RNN were explored to predict the direction of daily stock price changes. Since stocks in the same industry sector are highly correlated, we propose to replace individual stock with their corresponding Exchange Traded Sector Funds (ETFs) and S&P 500. We show that this replacement of individual stocks by the corresponding ETF can be justified due to high degree of returns correlation and small Hamming distances of trading signals across both input and output for stocks in the same sector. We considered historical daily data spanning 25 years (2000 to 2024) on major sector ETFs, some of their components and the S&P-500 index. Our results show that LSTM consistently outperformed CNN and RNN and traditional machine learning models. As a trading strategy, LSTM significantly out-performed buy-and-hold strategy for stocks in the Technology and Consumer Durables sectors, but offered very minor improvement for stocks in other sectors. Our results suggest that predicting daily stock price directions with deep learning models should not be used for most stocks unless these stocks are in Technology or Consumer Durables sectors and LSTM-based models are used.
Keywords: Deep learning; Stock movement prediction; Trading strategies; Time-series analysis