Domain-adapted Learning and Interpretability: DRL for Gas Trading

Paper: arXiv 2301.08359

Authors: Yuanrong Wang, Yinsen Miao, Alexander CY Wong, Nikita P Granger, Christian Michler

Abstract

Deep Reinforcement Learning (Deep RL) has been explored for a number of applications in finance and stock trading. In this paper, we present a practical implementation of Deep RL for trading natural gas futures contracts. The Sharpe Ratio obtained exceeds benchmarks given by trend following and mean reversion strategies as well as results reported in literature. Moreover, we propose a simple but effective ensemble learning scheme for trading, which significantly improves performance through enhanced model stability and robustness as well as lower turnover and hence lower transaction cost. We discuss the resulting Deep RL strategy in terms of model explainability, trading frequency and risk measures.

Complexity vs Empirical Score

• Math Complexity: 5.5/10
• Empirical Rigor: 7.5/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper provides a practical and novel application of Deep Reinforcement Learning to natural gas futures trading, addressing key real-world challenges like transaction costs and stability via an ensemble scheme. It balances formal DRL formulations with strong empirical backtesting against standard industry benchmarks, alongside valuable insights into model explainability.

Research Flowchart

  flowchart TD
    A["Research Goal: Optimize Natural Gas Futures Trading via DRL"] --> B["Data Inputs: Historical Natural Gas Futures & Market Features"]
    B --> C["Methodology: Train Deep RL Agents vs. Baseline Strategies"]
    C --> D["Ensemble Scheme: Combine DRL Models for Stability & Low Turnover"]
    D --> E["Evaluation: Assess Sharpe Ratio, Risk Measures & Model Interpretability"]
    E --> F["Outcomes: Superior Sharpe Ratio, Reduced Transaction Costs & High Stability"]