Gated deep neural networks for implied volatility surfaces

Yu Zheng, Yongxin Yang, Bowei Chen

Abstract

This study presents a novel framework for predicting implied volatility (IV) surfaces using gated deep neural networks. By integrating financial domain knowledge—such as no-static arbitrage, limiting boundaries, and the characteristic "volatility smile"—directly into the network architecture and loss function, the model achieves high predictive accuracy while maintaining mathematical consistency. Evaluated on 20 years of S&P 500 index option data, the proposed multi-model architecture outperforms traditional surface stochastic volatility inspired (SSVI) models and standard neural networks in both in-sample and out-of-sample tests.

Introduction: The Challenge of Implied Volatility

In financial markets, implied volatility represents the market's consensus on future asset price fluctuations. When plotted against an option's strike price (moneyness) and time to maturity, it forms an "implied volatility surface." Accurately modeling this surface is critical for pricing and risk management. However, researchers face a persistent trade-off: traditional parametric models like SSVI are mathematically robust but often too rigid to capture complex market dynamics, while standard "black box" neural networks are flexible but frequently produce surfaces that violate fundamental financial laws, such as the absence of arbitrage.

Methodology: Gated Multi-Model Architecture

The core innovation of this research is the Gated Multi-Model Architecture. Rather than relying on a single monolithic network, the system employs a "mixture of experts" approach.

The gated multi-model architecture combines several single models (a) into a comprehensive multi-model (b) using a gating network to weight contributions.

As shown in the architecture diagram above, a gating network analyzes the input parameters (moneyness $m$ and time to maturity $\tau$) to determine the optimal weights for several underlying "single models." This allows the network to specialize its learning across different regions of the volatility surface—for instance, handling short-term weekly options differently than long-term leaps.

Incorporating Financial Constraints

To ensure the model remains "rationally" constrained, the authors implemented several key features:

  • Custom Activation Function: A new "smile function" $\phi(z)$ based on hyperbolic tangents was developed to naturally reproduce the U-shaped volatility smile while remaining twice differentiable.
  • Constrained Optimization: The training loss function includes five specific penalty terms that punish violations of financial properties like monotonicity and the absence of butterfly arbitrage.
  • Smoothness: By ensuring the model is twice differentiable, it remains practical for calculating second-order risk sensitivities (Greeks) like Gamma.

Data and Empirical Analysis

The model was trained and validated using a massive dataset of S&P 500 index options from OptionMetrics, spanning over two decades from January 1996 to April 2016.

Summary of the 20-year S&P 500 dataset, showing (a) number of contracts, (b) number of quotes, (c) time to maturity ranges, and (d) log moneyness bounds over time.

The dataset includes approximately 2.9 million valid quotes across 63,338 option contracts. As illustrated in the data summary, the market grew significantly in complexity and volume over the study period, particularly after 2008. To ensure the model remained robust in regions with sparse market data, the authors supplemented the real-world quotes with synthetic samples to enforce constraints across the entire domain.

Key Results

The Multi-model was benchmarked against the industry-standard SSVI model and vanilla neural networks. Key findings include:

  • Superior Accuracy: The Multi-model achieved a Mean Average Percentage Error (MAPE) of 3.34% on out-of-sample data, significantly outperforming the SSVI benchmark (3.73%) and the Vanilla NN (4.46%).
  • Constraint Adherence: The model maintained a violation rate of less than 0.1% for financial constraints, proving that the penalty-based training effectively "tamed" the neural network's flexibility.
  • Stability: The architecture demonstrated resilience across various market regimes, including the 2008 financial crisis, maintaining stable performance where simpler models often failed.

Conclusion

This research demonstrates that machine learning performance does not need to be sacrificed for financial validity. By seamlessly combining data-driven gated architectures with domain-specific mathematical constraints, the authors provide a robust tool for implied volatility surface prediction. Future work may extend this framework to model the dynamic evolution of surfaces over time (stochastic volatility) and test its performance on less liquid asset classes.

Summary updated

Abstract

This paper presents a framework of developing neural networks to predict implied volatility surfaces. It can incorporate the related properties from existing mathematical models and empirical findings, including no static arbitrage, limiting boundaries, asymptotic slope and volatility smile. These properties are also satisfied empirically in our experiments with the option data on the S&P 500 index over 20 years. The developed neural network model outperforms the widely used surface stochastic volatility inspired (SSVI) model and other benchmarked neural network models on the mean average percentage error in both in-sample and out-of-sample datasets. This study has two major contributions. First, it contributes to the recent use of machine learning in finance, and an accurate deep learning implied volatility surface prediction model is obtained. Second, it provides the methodological guidance on how to seamlessly combine data-driven models with domain knowledge in the development of machine learning applications.
BibTeX
@inproceedings{zheng2021gated,
  author = {Yu Zheng and Yongxin Yang and Bowei Chen},
  title = {{Gated deep neural networks for implied volatility surfaces}},
  booktitle = {ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year = {2021},
  doi = {10.1145/3447548.3467115},
}
Figure