Abstract
We present a mixture of normalizing flows (MoNF) approach to European option pricing with guarantees that its estimations are free from static arbitrage. In contrast to many existing methods that meet economic rationality constraints (e.g., non-arbitrage) by introducing auxiliary losses, our solution meets those constraints exactly by design. To achieve this, we propose to build a model for risk neutral density using normalizing flows, which results in a pricing model, instead of modelling the option pricing function directly. First, we convert the constraints for direct pricing models to the constraints for models backed by risk neutral density estimation, then we design a specific NF architecture that meets these constraints. Furthermore, we find that employing a mixture of such normalizing flows improves the performance significantly, compared to using a deeper single NF. Finally, we present a mechanism to regularise the proposed model, and this regularisation can serve as a bridge between our method and any sample-based mathematical finance method. The evaluations on five option datasets show superiority of our method compared to mathematical finance solutions and some other neural networks based methods.
Introduction
Option pricing in modern finance is often a trade-off between the rigorous, "no-arbitrage" guarantees of Mathematical Finance (MF) models and the high-capacity data-fitting capabilities of Machine Learning (ML). While MF models like Black-Scholes are economically sound, they often fail to capture the complexities of real-world market data. Conversely, standard ML models can fit data with high precision but often violate fundamental economic principles—for instance, predicting that an option's price will increase as its strike price increases, which creates a risk-free profit (arbitrage) opportunity.
This paper introduces the Mixture of Normalizing Flows (MoNF), a framework that achieves the flexibility of deep learning while strictly adhering to the laws of financial economics. By modeling the underlying asset's risk-neutral density (RND) rather than the option price directly, the MoNF model is mathematically incapable of producing arbitrage opportunities.
Methodology: Constraints by Design
The core innovation lies in shifting the modeling target. Instead of training a neural network to output a price, the authors train it to model the Risk-Neutral Density (RND)— the probability distribution of the asset's price at expiration.
Normalizing Flows and Mixtures
The model uses Normalizing Flows (NF) to transform a simple Gaussian distribution into a complex, multi-modal density through invertible mappings. To ensure numerical stability and avoid vanishing gradients common in deep architectures, the authors employ a Mixture Model (MoNF). Rather than one very deep network, they use a mixture of several shallower NFs, finding that "width" (multiple experts) provides better performance for financial densities than "depth."
Wasserstein Regularization
To prevent the model from overfitting or becoming "too wiggly," a regularization term based on the Wasserstein distance (Earth Mover's Distance) is introduced. This pulls the learned density toward a simpler, well-behaved reference model, effectively bridging the gap between powerful ML flexibility and trusted MF simplicity.
Key Insights and Results
The MoNF model was evaluated on a massive dataset of 35 million option contracts (including Alphabet, Apple, Microsoft, NASDAQ100, and S&P500) from 2017 to 2021.
Performance and Robustness
MoNF consistently outperformed traditional MF models (like Heston or Variance Gamma) and existing constrained neural network baselines. In S&P500 interpolation tasks, MoNF achieved a Mean Absolute Error (MAE) of 0.0310, significantly lower than Black-Scholes (0.5354). Crucially, in extrapolation tasks—where standard ML models often produce nonsensical results—MoNF remained stable and economically rational.
The Power of Mixtures
A key finding was the superiority of mixture architectures. As shown in the experimental results, a mixture of shallower flows (e.g., 8 flows of 4 layers each) consistently achieved lower testing MAE than a single deep flow with the same total number of layers (e.g., 32 layers).
Conclusion
The MoNF approach demonstrates that economic constraints can be integrated into deep learning models by design rather than through penalty functions. By modeling the risk-neutral density, the framework ensures arbitrage-free pricing while maintaining high predictive accuracy. Future work aims to extend this from pricing curves (strike-based) to full pricing surfaces (incorporating time to maturity) while maintaining "no-calendar-arbitrage" constraints.
Summary updated