In quantitative finance, the "no-arbitrage" principle is a fundamental law: it should be impossible to make a guaranteed profit with zero risk. Traditional econometric models are built on this bedrock, but they often struggle to capture the messy, non-linear complexities of real-world market data. Conversely, modern deep learning models are excellent at fitting data but are "economically blind." Without explicit guidance, a standard neural network might predict option prices that violate basic logic—such as suggesting that an option with a higher strike price is worth more than one with a lower strike.
To bridge this gap, researchers are developing "rational by design" architectures that bake economic axioms directly into the machine learning framework.
The Cost of Unconstrained Learning
Standard machine learning models prioritize statistical fit above all else. As Jang et al. et al. observe, this often leads to price predictions that permit risk-free profit opportunities, rendering the models unreliable for professional trading or risk management. Dierckx and Schoutens point out that while purely data-driven architectures like Variational Autoencoders can generate statistically plausible volatility surfaces, they lack the mathematical safeguards to ensure those surfaces are financially consistent.
When a model predicts a surface that allows for arbitrage, it isn't just a minor error; it is a violation of the "physics" of the financial market. Ackerer et al. argue that this lack of inductive bias makes black-box models practically invalid for high-stakes applications where mathematical consistency is mandatory.
Rationality by Design
One way to solve this is to change what the network learns. Instead of asking a network to predict an option price directly, Yang et al. (2023) propose modeling the underlying asset's risk-neutral density using a Mixture of Normalizing Flows. Because the price is then derived by integrating over this density, the model is mathematically incapable of producing arbitrage opportunities. It is "rational by construction."
Similarly, Yang et al. (2017) introduced gated neural network architectures that enforce economic axioms—like the requirement that call option prices must be monotonic and convex relative to the strike price—directly through the network's internal structure. By embedding these laws into the architecture, the authors ensure the model provides a valid risk-neutral density while outperforming traditional econometric benchmarks.
Enforcing Consistency through Constraints
If the architecture itself cannot be restricted, the training process can be. This involves adding "soft" or "hard" constraints to the loss function that penalize the model whenever it strays into economically impossible territory.
- Penalty Functions: Zheng et al. utilize a custom "smile" activation function combined with a loss function that explicitly penalizes arbitrage violations.
- Auxiliary Tasks: Pelger demonstrates that incorporating the no-arbitrage condition as a constraint helps the model disentangle true risk signals from market noise, significantly improving the estimation of the Stochastic Discount Factor.
- Real-time Smoothing: For high-frequency environments, Yang et al. (2025) use hypernetworks to generate the parameters of a volatility surface in milliseconds. By using auxiliary loss functions during training, they ensure these surfaces remain arbitrage-free even when data is sparse.
From Points to Operators
The latest evolution in this field moves away from predicting individual prices toward learning the entire "operator" that governs the market surface. Wiedemann et al. introduce Operator Deep Smoothing, which uses Graph Neural Operators to map irregular market observations to a continuous, smooth, and arbitrage-free implied volatility surface.
This shift treats the no-arbitrage condition not just as a rule to be followed, but as a structural property of the financial environment. By treating these economic laws as the "laws of physics" for the model, researchers are creating AI systems that are both as flexible as deep learning and as rigorous as classical finance.
Key Foundational Work
A general framework for building machine learning models for pricing american index options with no-arbitrage and its limitation (2016) — contributed 5 concept elements to this direction.
Gated Neural Networks for Option Pricing: Rationality by Design (2017) — contributed 1 concept elements to this direction.