Scholarlink

<p>This research investigates wavelet-enhanced deep reinforcement learning (DRL) for trading S&amp;P&#xa0;500 futures, assessing four wavelet families (Daubechies, Symlets, Coiflets, and Biorthogonal) in conjunction with three DRL algorithms (PPO, A2C, and DQN). We employ level-2 decomposition utilizing conservative soft thresholding on market microstructure indicators (DIX, GEX, VIX), enhancing signal-to-noise ratios by 25–41 dB. The <Emphasis FontCategory="NonProportional">coif4+DQN</Emphasis> combination yields the most robust outcomes (Sharpe ratio: 0.96; total return: 112.5%), while <Emphasis FontCategory="NonProportional">A2C+db1</Emphasis> and <Emphasis FontCategory="NonProportional">PPO+db4</Emphasis> attain Sharpe ratios of 0.803 and 0.801, respectively. In comparison with XGBoost, random forest, and Logistic regression, wavelet-DRL attains a 35–70% superior Sharpe ratio and reduces maximum drawdowns (0.28–0.34 vs. 0.39–0.42). Feature-set analysis reveals that DIX, GEX, and VIX combined surpass single-indicator configurations by 18.7% in Sharpe ratio. Statistical analyses validate the robustness, revealing that 82.1% of maximum drawdown (MDD) and 67.9% of Sharpe ratio comparisons are significant at the 5% level. Our findings support a gradual implementation strategy—transitioning from <Emphasis FontCategory="NonProportional">A2C+db1</Emphasis> to <Emphasis FontCategory="NonProportional">DQN+coif4</Emphasis>—to enhance institutional algorithmic trading efficacy.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing algorithmic trading with wavelet-based deep reinforcement learning: a multi-indicator approach

  • 期刊论文

  • 发表日期：

    2025年9月14日

查看全文
  • Antonio José Martínez Casares

摘要

This research investigates wavelet-enhanced deep reinforcement learning (DRL) for trading S&P 500 futures, assessing four wavelet families (Daubechies, Symlets, Coiflets, and Biorthogonal) in conjunction with three DRL algorithms (PPO, A2C, and DQN). We employ level-2 decomposition utilizing conservative soft thresholding on market microstructure indicators (DIX, GEX, VIX), enhancing signal-to-noise ratios by 25–41 dB. The coif4+DQN combination yields the most robust outcomes (Sharpe ratio: 0.96; total return: 112.5%), while A2C+db1 and PPO+db4 attain Sharpe ratios of 0.803 and 0.801, respectively. In comparison with XGBoost, random forest, and Logistic regression, wavelet-DRL attains a 35–70% superior Sharpe ratio and reduces maximum drawdowns (0.28–0.34 vs. 0.39–0.42). Feature-set analysis reveals that DIX, GEX, and VIX combined surpass single-indicator configurations by 18.7% in Sharpe ratio. Statistical analyses validate the robustness, revealing that 82.1% of maximum drawdown (MDD) and 67.9% of Sharpe ratio comparisons are significant at the 5% level. Our findings support a gradual implementation strategy—transitioning from A2C+db1 to DQN+coif4—to enhance institutional algorithmic trading efficacy.