In the physical sciences, models are bound by the laws of thermodynamics and gravity. In finance, the equivalent "laws of physics" are no-arbitrage conditions—the principle that market prices should not allow for risk-free profit. While deep learning has revolutionized predictive modeling, standard "black-box" neural networks are notoriously indifferent to these rules. A model might achieve high statistical accuracy on historical data while simultaneously predicting an implied volatility surface that is mathematically impossible, rendering it useless for actual trading or risk management.
A growing body of research seeks to bridge this gap by embedding economic axioms directly into machine learning architectures. The goal is to create models that are "rational by design," combining the high-capacity pattern recognition of deep learning with the structural rigor of classical mathematical finance.
The Cost of Unconstrained Flexibility
Standard neural networks are universal function approximators, but in finance, this flexibility is a double-edged sword. As Damien Ackerer, Natasa Tagasovska, and Thibault Vatter argue in Deep Smoothing of the Implied Volatility Surface, a lack of inductive bias allows models to explore unrealistic market states. For example, an unconstrained model might predict that an option with a higher strike price is more expensive than one with a lower strike—a clear violation of monotonicity that creates a "free money" opportunity.
These violations aren't just theoretical nuisances; they lead to numerical instability and poor generalization. Yu Zheng and Yongxin Yang note in Incorporating Prior Financial Domain Knowledge into Neural Networks for Implied Volatility Surface Prediction that purely data-driven models often fail to respect the "smile" shape of volatility, making them unreliable for professional pricing environments where mathematical consistency is mandatory.
Soft Constraints via Arbitrage-Free Regularization
The most common way to enforce financial laws is through the loss function. By adding penalty terms that trigger when the model violates economic principles—such as strike price convexity or time-to-maturity monotonicity—researchers can "nudge" the network toward valid states.
In Risk-Neutral Generative Networks, Zhonghao Xian and colleagues demonstrate how formulating no-arbitrage conditions as penalty terms ensures that generative models produce valid risk-neutral densities. Similarly, Cody Hyndman shows in Deep Learning in a Generalized HJM-type Framework Through Arbitrage-Free Regularization that penalizing "free money" opportunities allows deep learning to model bond markets while maintaining the theoretical guarantees of the Heath-Jarrow-Morton framework.
This "soft" approach is highly flexible. For instance, the HyperIV framework by Yongxin Yang and Timothy M. Hospedales uses auxiliary loss functions to generate entire arbitrage-free volatility surfaces in milliseconds, making it practical for high-frequency trading where speed and validity must coexist.
Hard Constraints and Rationality by Design
While regularization "encourages" validity, it does not "guarantee" it. To achieve strict adherence to financial laws, researchers have developed architectures where the constraints are baked into the network's very structure.
In Gated Neural Networks for Option Pricing: Rationality by Design, Yongxin Yang and Timothy Hospedales introduce a "divide-and-conquer" gated architecture. By using specialized activation functions and architectural bottlenecks, the model is mathematically incapable of producing prices that violate no-arbitrage axioms. This "rationality by design" ensures that the model's output always represents a valid risk-neutral density, even when extrapolating into regions with sparse data.
Hybrid Stochastic Frameworks
Another direction involves merging deep learning with the formal language of stochastic calculus. Rather than replacing traditional models, neural networks are used to parameterize them.
Patryk Gierjatowicz and Lukasz Szpruch explore this in Robust pricing and hedging via neural SDEs. By using neural networks to define the drift and diffusion coefficients of a Stochastic Differential Equation (SDE), they ensure the discounted asset price remains a local martingale. This structural choice guarantees no-arbitrage by construction, providing a data-driven way to calculate robust price intervals for complex derivatives.
Beyond Options: Economic Physics in Other Domains
The principle of "physics-informed" finance extends beyond option pricing to any domain governed by economic laws.
- Demand Forecasting: In PREIG: Physics-informed and Reinforcement-driven Interpretable GRU for Commodity Demand Forecasting, Hongwei Ma and Junbin Gao embed the principle of negative price elasticity into a GRU's loss function. This prevents the model from making the "economically illiterate" prediction that increasing prices will increase demand.
- Urban Planning: Haohao Qu and Jun Li apply similar logic to electric vehicle charging demand in PAG, ensuring that predictions remain consistent with common-sense economic behavior even when training data is noisy.
- Combinatorial Auctions: Jakob Heiss and Sven Seuken utilize Monotone-Value Neural Networks to ensure that adding items to a bundle never decreases its predicted value, a fundamental requirement for rational agent behavior in spectrum auctions.
By treating economic principles as hard constraints rather than optional suggestions, this research direction transforms neural networks from fragile curve-fitters into robust tools capable of navigating the complex, rule-bound reality of global markets.