Deep Reinforcement Learning for Quantitative Trading: A Novel Framework Integrating Variational Mode Decomposition and Contrastive Transformer

Wensheng Yi, Gao Ke, Jiaming Chen

Abstract

To address the limitations of traditional quantitative trading models in extracting non-linear features, modeling long-term dependencies, and identifying abnormal fluctuations in financial time series, this paper proposes an improved quantitative trading framework integrating Variational Mode Decomposition (VMD), self-supervised Contrastive Learning, and Deep Reinforcement Learning (DRL). This framework aims to resolve the challenges of low Signal-to-Noise Ratio (SNR) and non-stationary distributions in high-frequency financial data. First, VMD is utilized to decompose raw price-volume data into several Intrinsic Mode Functions (IMFs), effectively separating market microstructure noise from core trend signals. Second, a Contrastive Transformer Encoder based on a time-aware mechanism is constructed to extract robust market state representations via self-supervised contrastive learning tasks, overcoming the overfitting issues inherent in supervised learning during extreme market conditions. Finally, an improved Proximal Policy Optimization (PPO) algorithm is introduced at the decision optimization layer, employing the Differential Sharpe Ratio as the core reward function to achieve end-to-end portfolio optimization. Backtesting results on the S&P 500 index (2024–2025) demonstrate that the proposed model significantly outperforms traditional LSTM and Mean-Variance models. Specifically, the model achieved an Annualized Rate of Return (ARR) of 28.45% and a Sharpe Ratio of 2.18, while maintaining a Maximum Drawdown of -9.15% during global market volatility, validating the potential of generative AI technologies in quantitative investment.

Source: semanticscholar · PDF

Read the AI summary, key takeaways and discussion on WOBR Quant Research.