A Deep Reinforcement Learning Approach to Automated Stock Trading, using xLSTM Networks (2503.09655v1)

Abstract: Traditional Long Short-Term Memory (LSTM) networks are effective for handling sequential data but have limitations such as gradient vanishing and difficulty in capturing long-term dependencies, which can impact their performance in dynamic and risky environments like stock trading. To address these limitations, this study explores the usage of the newly introduced Extended Long Short Term Memory (xLSTM) network in combination with a deep reinforcement learning (DRL) approach for automated stock trading. Our proposed method utilizes xLSTM networks in both actor and critic components, enabling effective handling of time series data and dynamic market environments. Proximal Policy Optimization (PPO), with its ability to balance exploration and exploitation, is employed to optimize the trading strategy. Experiments were conducted using financial data from major tech companies over a comprehensive timeline, demonstrating that the xLSTM-based model outperforms LSTM-based methods in key trading evaluation metrics, including cumulative return, average profitability per trade, maximum earning rate, maximum pullback, and Sharpe ratio. These findings mark the potential of xLSTM for enhancing DRL-based stock trading systems.

• The paper introduces a deep reinforcement learning approach for automated stock trading that leverages xLSTM networks, an architecture designed to better capture long-term dependencies and manage gradients than traditional LSTMs.
• Empirical results show the xLSTM-based DRL system achieves significantly higher cumulative returns and average profitability per trade compared to standard LSTM models on historical market data.
• The xLSTM architecture also demonstrates improved risk management, exhibiting a better maximum pullback and a marked improvement in the Sharpe ratio, indicating superior risk-adjusted returns.

Introduction

The paper "A Deep Reinforcement Learning Approach to Automated Stock Trading, using xLSTM Networks" (2503.09655) explores an advanced integration of extended LSTM (xLSTM) architectures with deep reinforcement learning (DRL) frameworks to enhance algorithmic trading strategies. The work is motivated by the inherent limitations of traditional LSTM architectures in capturing long-term dependencies and mitigating the vanishing gradient problem, both of which are critical in processing volatile financial time-series data.

Methodology

The paper introduces an xLSTM variant that incorporates exponential gating mechanisms alongside a restructured memory design comprising sLSTM and mLSTM blocks. The former aims to improve gradient flow and convergence, while the latter provides enhanced parallel processing and memory management, which are pivotal for DRL applications. The xLSTM networks are employed within both the actor and critic modules of the DRL framework. The selection of Proximal Policy Optimization (PPO) as the DRL algorithm underscores the method’s emphasis on achieving an optimal trade-off between exploration and exploitation during the training phase.

Experimental Setup and Results

Empirical evaluations were conducted using historical market data from key technology firms, including NVIDIA, Apple, Microsoft, Google, and Amazon over an extensive time horizon. The experimental framework benchmarks the xLSTM-based DRL system against standard LSTM-based models across several performance metrics:

• Cumulative Returns: The xLSTM model achieved significantly higher cumulative returns.
• Average Profitability per Trade: Enhanced profit per trade was observed with xLSTM, indicating improved trade selection and execution.
• Maximum Earning Rate: The proposed model exhibited superior maximum earnings, suggesting better capture of high-profit opportunities.
• Maximum Pullback: The architecture managed drawdowns more efficiently, reflecting a robust risk management profile.
• Sharpe Ratio: A marked improvement in risk-adjusted returns was noted, as seen by elevated Sharpe ratios relative to baseline LSTM models.

These quantitative improvements firmly establish that the architectural enhancements in xLSTM translate into more effective financial decision-making under dynamic market conditions.

Discussion and Future Work

The empirical results support the strong claim that leveraging xLSTM architectures substantially enhances the performance of DRL-based trading systems. Notably, the improvements in both directional profitability and risk-adjusted metrics (such as the Sharpe ratio) underscore the potential of xLSTM networks for handling the non-stationarity and high volatility of stock market data. However, the paper acknowledges the increased computational overhead associated with training xLSTM networks, indicating a trade-off between performance gains and computational efficiency. Future research may focus on optimizing the training process, integrating advanced feature engineering, and exploring ensemble modeling approaches to enhance scalability and robustness further.

Conclusion

The paper presents a comprehensive and technically rigorous approach that integrates extended LSTM structures into a DRL framework using PPO for automated stock trading. By addressing critical deficiencies of classical LSTM models in handling long-range dependencies and gradient issues, the xLSTM-based architecture demonstrates significant performance enhancements across key trading metrics including cumulative returns, average trade profitability, maximum earning rate, maximum pullback, and Sharpe ratio. In summary, the paper provides a compelling case for the adoption of xLSTM networks in DRL applications tailored for financial markets, while also outlining the necessity for further research to manage computational complexities.

In summary, the paper develops an advanced DRL trading strategy leveraging xLSTM networks that deliver robust performance improvements, while also highlighting computational challenges that warrant future investigation.