Paper: arXiv 2609.33767
Authors: Wee Ling Tan, Stephen Roberts, Stefan Zohren
Abstract
We present an end-to-end deep learning framework for systematic options trading that directly embeds hedging behavior through explicit control of portfolio-level risk exposures. While neural networks trained to optimize risk-adjusted performance have been shown to outperform traditional rules-based strategies, such approaches remain agnostic to the sensitivities of the resulting portfolios with respect to specific underlying risk factors. We propose a general training objective that combines a performance-driven loss with a differentiable risk-sensitivity penalty, enforcing neutrality to selected risk dimensions. Unlike reinforcement learning methods that approximate optimal hedging policies via simulated market dynamics, our framework operates entirely on historical data and jointly optimizes risk-adjusted returns and targeted risk constraints in a single learning problem. We instantiate the framework on static delta-neutral straddle portfolios with the penalty directed at first-order directional exposure, and evaluate two penalty variants – an exposure-normalized penalty and a Greek-ratio drift penalty. Empirical results on Nasdaq 100 equity options demonstrate that appropriately calibrated regularization simultaneously improves out-of-sample risk-adjusted performance relative to an unregularized baseline while reducing realized directional exposure.
Complexity vs Empirical Score
- Math Complexity: 6.5/10
- Empirical Rigor: 8.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: This paper presents a novel deep learning framework for options trading that integrates risk-sensitivity penalties directly into the learning objective, moving beyond traditional risk-agnostic approaches. The empirical evaluation on Nasdaq 100 equity options demonstrates strong out-of-sample performance and effective risk control, indicating a robust methodology. The combination of advanced mathematical concepts in deep learning and rigorous empirical validation places it firmly in the ‘Holy Grail’ quadrant.
Research Flowchart
flowchart TD
A["Research Goal: Tame Greeks in Option Portfolios"] --> B{"Key Problem: Traditional NN agnostic to risk sensitivities"};
B --> C["Methodology: Deep Learning Framework with Differentiable Risk-Sensitivity Penalty"];
C --> D["Inputs: Historical Options Data (Nasdaq 100 Equities)"];
D --> E["Computational Process: Joint Optimization of Performance & Targeted Risk (Delta-Neutral Straddle Portfolios)"];
E --> F{"Outcomes: Improved Risk-Adjusted Performance & Reduced Realized Directional Exposure"};
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes an end-to-end deep learning framework that combines a performance-driven loss with a differentiable risk‑sensitivity penalty, introducing exposure‑normalized and Greek‑ratio drift penalty variants.
Source sentence
We propose a hybrid training objective L(θ) that combines a performance-driven objective with an explicit risk-sensitivity penalty.
Assumes 5
- Static delta‑neutral hedging is performed only at initiation with no subsequent adjustments. execution
Source sentence
This static hedging approach is consistent with [13, 15, 31], where a one-time delta hedge is implemented only at initiation, with no subsequent adjustments to the weights.
- Option Greeks and implied volatilities are computed using a binomial tree model (Cox‑Ross‑Rubinstein). model
Source sentence
The IVs and Greeks are calculated with a binomial tree model using Cox, Ross, and Rubinstein [8].
- European‑style execution is assumed for near‑ATM options. market
Source sentence
This approach to computing returns inherently assumes European-style execution, given our focus on near-ATM options that retain substantial extrinsic value.
- Transaction costs are proportional to daily turnover measured in basis points. other
Source sentence
where c denotes the proportional cost per unit of turnover, expressed in basis points.
Show 1 more
- Liquidity is sufficient; the dataset comprises deep and liquid options markets for Nasdaq‑100 constituents. market
Source sentence
This universe comprises a broad cross-section of large-cap US equities and are characterized by relatively deep and liquid options markets.
Data
Nasdaq 100 equity options · options · daily
- OptionMetrics Ivy DB options commercial
Source sentence
... we restrict our analysis to the universe of option contracts written on constituents of the Nasdaq 100 Index over the period from January 2010 to December 2023.
Costs
Modeled
Costs are incorporated as a proportional fee c (in basis points) applied to daily turnover in the turnover‑adjusted return calculation.
Source sentence
where c denotes the proportional cost per unit of turnover, expressed in basis points.
Finds 5
Sharpe 2.34 (benchmark 0.62) · max drawdown 21.0%
- Regularized models improve out‑of‑sample risk‑adjusted performance relative to an unregularized baseline while reducing realized directional exposure. performance
Source sentence
appropriately calibrated models not only exhibit competitive out-of-sample hedging performance but also simultaneously improve risk-adjusted returns relative to an unregularized baseline.
- DP (L1) achieves the highest Sharpe ratio of 2.339 at zero transaction cost, outperforming all other variants. performance
Source sentence
DP (L1) achieving the highest Sharpe ratio of 2.339.
- At high transaction costs (100 bps), the baseline LSTM with turnover regularization attains the top Sharpe (0.871), slightly ahead of DP (L1) at 0.860. performance
Source sentence
At the highest cost level of 100 bps, Baseline LSTM + TC Reg achieves the highest Sharpe ratio of 0.871, followed closely by DP (L1) at 0.860.
- ENP (L2) retains a Sharpe ratio above 2.0 across regularization strengths α = 100‑104 while reducing median net Delta to 0.072. performance
Source sentence
ENP (L2), in particular, retains a Sharpe ratio above 2.0 across α = 100 through 104, while eventually reducing the median net Delta to 0.072.
- Increasing regularization beyond moderate levels sharply deteriorates performance as the model’s capacity to generate signals is suppressed. negative
Source sentence
The deterioration in performance observed at extreme values of α confirms that hedging is not costless: once the risk-sensitivity penalty is given overwhelming weight, the model’s capacity to generate meaningful trading signals is suppressed.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 57 related papers). Coverage is partial, so this is neither replication nor refutation.
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (49)
- Time Series Momentum (2012) — link
- Dissecting Investment Strategies in the Cross Section and Time Series (2015)
- Spatio-Temporal Momentum : Jointly Learning Time-Series and Cross-Sectional Strategies — → 2302.10175
- Enhancing Time-Series Momentum Strategies Using Deep Neural Networks (2019) — link
- Slow Momentum with Fast Reversion : A Trading Strategy Using Deep Learning and Changepoint Detection (2022) — link
- Trading with the Momentum Transformer : An Intelligent and Interpretable Architecture (2023) — link
- Returns to Buying Winners and Selling Losers : Implications for Stock Market Efficiency (1993) — link
- International Momentum Strategies (1998)
and 41 more
- Momentum in Futures Markets (2005)
- Building Cross-Sectional Systematic Strategies by Learning to Rank (2021) — link
- Enhancing Cross-Sectional Currency Strategies by Context-Aware Learning to Rank with Self-Attention (2022) — link
- Mean Reversion in Stock Prices : Evidence and Implications (1988) — link
- Does the Stock Market Overreact ? (1985) — link
- Fact and Fantasy in the Use of Options (1975) — link
- OCC - Historical Volume Statistics (2024)
- Deep Learning for Options Trading: An End-To-End Approach — → 2407.21791
- The Pricing of Options and Corporate Liabilities (1973)
- Theory of Rational Option Pricing (1973) — link
- Option Pricing : A Simplified Approach (1979) — link
- A Factor Model for Option Returns (2022) — link
- Maturity Driven Mispricing of Options (2022) — link
- Option Pricing and Replication with Transactions Costs (1985)
- Option Replication in Discrete Time with Transaction Costs (1992)
- Options Arbitrage in Imperfect Markets (1989)
- Deep Hedging (2019) — link
- Dynamic Replication and Hedging : A Reinforcement Learning Approach (2019) — link
- Deeper Hedging: A New Agent-based Model for Effective Deep Hedging — → 2310.18755
- Adversarial deep hedging: Learning to hedge without price process modeling — → 2307.13217
- Fast deep hedging with second-order optimization — → 2410.22568
- A Nonparametric Approach to Pricing and Hedging Derivative Securities Via Learning Networks (1994) — link
- Option Pricing using Machine Learning (2021) — link
- Option Return Predictability with Machine Learning and Big Data (2023) — link
- Expected Option Returns (2001) — link
- Cross-section of Option Returns and Volatility (2009) — link
- Equity Volatility Term Structures and the Cross Section of Option Returns (2017) — link
- Option Momentum (2023) — link
- Who Profits from Trading Options ? (2023) — link
- Limits of Arbitrage and Primary Risk-Taking in Derivative Securities (2023) — link
- Options Trading Costs Are Lower than You Think (2020) — link
- Options, futures, and other derivatives (2016)
- Approximation of Changes in Option Values and Hedge Ratios: How Large Are the Errors? (1997)
- The Sharpe Ratio (1994) — link
- The Impact of Volatility Targeting (2018) — link
- Long Short-Term Memory (1997) — link
- Time Series Forecasting With Deep Learning : A Survey (2021) — link
- Autoregressive Convolutional Neural Networks for Asynchronous Time Series (2018)
- Adam: A Method for Stochastic Optimization (2014) — link
- DeepLOB : Deep Convolutional Neural Networks for Limit Order Books (2019)
- Deep Learning for Portfolio Optimization (2020)
5 of 49 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.