What question did this study set out to answer?

To develop a hybrid neural network model that improves stock return predictions using high-frequency data.

June 14, 2026Open Access

Deep learning and stock returns: neural networks combining sequential information and cross-sectional information

Key Points

To develop a hybrid neural network model that improves stock return predictions using high-frequency data.
Proposed a novel LSTM-MLP hybrid architecture for predictive modeling.
Utilized high-frequency intraday data from the TAQ database for analysis.
Measured performance against Random Forest and XGBoost models in terms of turnover ratio and profitability.
The hybrid model achieved a turnover ratio of 43.7%, significantly lower than Random Forest (77.5%) and XGBoost (72.2%).
Demonstrated superior out-of-sample returns when adjusted for trading costs.
Captured unique return patterns not accounted for by existing models, providing valuable insights for portfolio management.

Abstract

The rise of advanced sensors, communication networks, and storage systems has facilitated the collection of high-frequency data across various sectors, including finance, telecommunications, and the Internet of Things (IoT). High-frequency data provides a greater variety of informational dimensions, thereby improving decision-making accuracy. However, the complex temporal and cross-sectional structures inherent in high-frequency data present significant challenges for conventional predictive modeling, particularly in financial markets. To overcome these limitations, we propose a novel hybrid neural network architecture that utilizes a structured, staged fusion of Long Short-Term Memory (LSTM) and Multi-Layer Perceptron (MLP) models. Unlike traditional ensemble methods that process features in parallel, our approach first extracts temporal embeddings via LSTM to preserve sequential dependencies and then integrates them with cross-sectional characteristics. We apply this LSTM-MLP hybrid model to the U.S. stock market, utilizing granular intraday transaction and quotation data from the Trade and Quote (TAQ) database to accurately estimate trading costs. Our empirical analysis demonstrates that this structure-preserving design significantly reduces signal volatility, resulting in a lower turnover ratio (43.7%) compared to Random Forest (77.5%) and XGBoost (72.2%). Consequently, the hybrid model outperforms both traditional hand-engineered strategies and state-of-the-art machine learning benchmarks in terms of out-of-sample, after-cost returns. Notably, the LSTM-MLP model achieves robust profitability while maintaining a lower turnover ratio compared to alternative approaches, highlighting its practical value in real-world trading environments. Regression diagnostics further reveal that the hybrid model captures distinctive return patterns not explained by existing models, thereby offering unique insights for portfolio management. Collectively, these findings advance the theoretical understanding of deep learning applications to high-frequency financial data and provide a robust, cost-aware decision support tool for practitioners. The proposed framework also holds promise for broader applications in other high-frequency data domains such as telecommunications and IoT.

Bookmark

View Full Paper