Deep Learning for Options Trading: An End-To-End Approach
Deep Learning for Options Trading: An End-To-End Approach
This paper presents a novel end-to-end deep learning framework for trading options, directly optimizing trading signals from market data without relying on traditional option pricing models or assumptions about market dynamics. The authors construct static delta-neutral straddle portfolios for S&P 100 stocks and train neural networks (MLP, LSTM, Transformer) to maximize the Sharpe ratio. Backtesting over 2010–2023 shows that deep learning models significantly outperform rules-based strategies like time-series momentum and MACD, achieving Sharpe ratios above 1.2 versus 0.8 for the best benchmark. Incorporating turnover regularization further enhances net returns by reducing transaction costs, with LSTM and Transformer models capturing temporal patterns most effectively. The framework is scalable, data-driven, and robust across market conditions, including the COVID-19 crisis. This work bridges the gap between machine learning and options trading, offering a practical alternative to traditional hedging and pricing approaches.
Highlights
- 1Introduces an end-to-end deep learning framework for options trading that directly learns trading signals from market data without requiring option pricing models or market dynamics assumptions.
- 2Demonstrates significant improvements in risk-adjusted performance over traditional rules-based strategies (e.g., TSMOM, MACD) on a backtest spanning 2010–2023 for S&P 100 options.
- 3Incorporates turnover regularization to enhance performance under high transaction costs, showing further gains in net returns.
- 4Provides a comprehensive comparison of neural network architectures (MLP, LSTM, Transformer) against momentum and mean-reversion benchmarks.
Methods
- MEnd-to-end neural network training: Optimizes trading signals directly using Sharpe ratio as the objective function, with architectures including MLP, LSTM, and Transformer.
- MVolatility targeting: Scales individual straddle returns by a 20-day exponentially weighted moving standard deviation to target 15% annualized volatility.
- MTurnover regularization: Adds a penalty on portfolio turnover to the loss function to reduce transaction costs.
- MStatic delta-neutral straddle construction: Forms portfolios of at-the-money straddles with one-time delta hedging at initiation.
Results
- RDeep learning models achieve annualized Sharpe ratios of 1.2–1.5, outperforming the best rules-based strategy (MACD momentum) with a Sharpe of 0.8.
- RTurnover regularization reduces average daily turnover from 0.6 to 0.3, improving net returns by 2–3% annually under realistic transaction costs.
- RLSTM and Transformer models show superior performance compared to MLP, capturing temporal dependencies in option returns.
- RAll deep learning models generate positive returns during the COVID-19 market selloff, while some benchmarks suffer losses.
- RThe end-to-end approach consistently outperforms both momentum and mean-reversion strategies across different market regimes.
Analyze Paper
Generate insights from "Deep Learning for Options Trading: An End-To-End Approach".