🧬 biology

Reliability Diagnostics for Physics-Informed Neural Option-Pricing Models: Residual-Error Geometry, Greek Stability, and No-Arbitrage Structure

This paper proposes a comprehensive a posteriori diagnostic framework that moves beyond simple residual loss to evaluate the reliability of Physics-Informed Neural Networks for option pricing by analyzing residual-error geometry, Greek stability, and no-arbitrage constraints, demonstrating through extensive benchmarks that these structural diagnostics provide a more accurate assessment of pricing errors than residual metrics alone.

Original authors: Francesco Demuro

Published 2026-07-16

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Francesco Demuro

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to teach a robot how to bake the perfect chocolate cake. In the world of finance, this "cake" is a financial contract called an option, which gives someone the right to buy or sell a stock at a specific price later on. To bake this cake correctly, the robot needs to follow a very strict recipe called a "partial differential equation" (PDE). Think of this PDE as the laws of physics for money; it dictates how the price of the option should change based on how the stock price moves, how much time is left, and how volatile the market is.

For a long time, mathematicians have used rigid, grid-based tools to solve these recipes, like a chef using a precise ruler to measure every ingredient. But recently, a new kind of robot has arrived: the Physics-Informed Neural Network (PINN). Instead of just following a recipe step-by-step, a PINN is like a super-smart apprentice that learns by tasting. It tries to bake the cake, checks if it violates the laws of physics (the PDE), and adjusts its recipe until the "taste test" (the error) is low. The big question for everyone in finance is: if this robot says the cake tastes perfect according to the laws of physics, can we trust it with real money? This paper dives into that exact question, exploring whether a low "taste error" actually means the robot has baked a reliable cake, or if it's just hiding a burnt bottom that no one noticed.

The Robot Chef and the "Perfect" Cake

In the world of high-stakes finance, getting the price of an option wrong isn't just a math error; it can mean losing millions of dollars. Financial engineers use these prices to decide how much to charge for insurance against stock crashes or how to hedge their bets. For years, they've relied on classical math methods to solve these pricing puzzles. But as markets get more complex, they've started turning to Artificial Intelligence, specifically a type of AI called Physics-Informed Neural Networks (PINNs).

Think of a PINN as a student who is being taught to solve a math problem not just by looking at the answer key, but by being forced to follow the rules of the universe. In finance, the "rules of the universe" are the no-arbitrage principles—basically, the idea that you can't get something for nothing. The AI is trained to minimize a "residual," which is a fancy word for the difference between what the AI predicts and what the laws of physics say should happen. If the residual is zero, the AI has perfectly followed the rules.

The big assumption many people have been making is simple: If the AI's residual is low (meaning it follows the rules well), then its price must be accurate. It's like assuming that if a student gets a perfect score on a multiple-choice quiz about the rules of baking, they must be able to bake a perfect cake.

The Reality Check: When Rules Don't Equal Results

Francesco Demuro, a researcher from Politecnico di Milano, decided to test this assumption. He set up a controlled experiment using the Black-Scholes model, which is the "textbook" way to price a simple European call option. He trained 60 different versions of these AI robots, tweaking their brains (architectures), their training seeds (random starts), and how much they cared about the rules versus the final taste (loss weighting).

He then compared these AI robots against two things:

The Gold Standard: The exact mathematical formula (the "perfect cake").
The Old School Method: A classical computer method called Crank-Nicolson, which is like a very precise, slow-moving ruler.

Here is the twist the paper reveals: Low residual loss is not a guarantee of a good price.

In the experiments, the AI robots that minimized their "rule-breaking" score (the residual) often still made significant mistakes in the actual price. Some of them were off by an average of 3.65 dollars (Mean Absolute Error), while a simple supervised neural network (one that just memorized the answer key) was off by only 0.19 dollars. Even more surprisingly, the AI robots that followed the rules perfectly sometimes produced prices that were "unstable."

Imagine a robot that calculates the price of a stock option. If the stock price moves up by a tiny bit, the robot should say the option price goes up a tiny bit. But in some of these AI models, a tiny move in the stock caused the option price to jump wildly or even go down when it should have gone up. This is like a robot chef who says, "If you add one more egg, the cake will be half the size!" That's a disaster for a hedge fund manager trying to protect their money.

The New Diagnostic Toolkit: Checking the Whole Cake

So, if the "residual score" isn't enough, how do we know if the AI is reliable? Demuro proposes a new "Reliability Score" that looks at the whole picture, not just the rules. He suggests checking four specific things, like a quality control inspector:

The Residual: Did it follow the laws of physics? (Yes, but that's not enough).
The Terminal Condition: Did it get the final taste right? At the very end of the option's life, the price must match the actual payoff (like a cake that must taste exactly like chocolate at the end). If the AI fails here, everything else is wrong.
No-Arbitrage Structure: Does the price behave logically? For example, if the stock price goes up, the option price must go up. If the AI says it goes down, it's violating the basic laws of money.
Greek Stability: Are the sensitivities (called "Greeks") stable? These are the numbers traders use to hedge. If the AI's "Delta" (how much the price changes with the stock) jumps around wildly, it's useless for risk management.

When Demuro combined all these checks into one "Full Reliability Score," he found something powerful. This new score was much better at predicting which AI models were actually accurate than just looking at the residual score alone. In fact, the correlation between this new score and the actual error was the strongest he found.

The Geometry of Mistakes

The paper also offers a fascinating insight into why this happens, using a concept called "Residual–Error Geometry." Imagine the error in the AI's prediction as a landscape. The "residual" is like a shadow cast by that landscape. Sometimes, the shadow looks small and neat, but the landscape itself has huge, hidden valleys or spikes that the shadow doesn't show.

The paper explains that for these financial equations, the residual is just a "differential image" of the error. This means the AI might be smoothing out the big mistakes in the price to make the "shadow" (the residual) look small, while the actual price is still wrong in specific, dangerous spots. It's like a student who memorizes the formula for a curve but draws the curve slightly wrong in the middle; the formula looks right, but the drawing is useless.

The Heston Pilot: A Bigger, Messier Kitchen

To see if this idea works in a more complex world, the author ran a small "pilot study" using the Heston model. This model is like a cake recipe that includes a second ingredient: random volatility (the "fear" of the market changing). This makes the math much harder and introduces "degenerate" areas where the math gets sticky.

Even in this pilot study with just two models, the "improved" version (which used the new diagnostic checks) performed better than the baseline. It had lower errors and a better reliability score. While this was just a small test, it suggests that the new diagnostic toolkit works even when the math gets messy and the "exact answer" is hard to find.

The Takeaway: Trust, But Verify

The main conclusion of this paper is a gentle but firm warning to the financial world: Do not trust a neural network just because it says it followed the rules.

In the controlled world of the Black-Scholes model, a supervised AI (one that just learned from the answer key) actually performed better than the physics-informed AI. But the real value of this paper isn't saying "AI is bad." It's saying that if you want to use AI in finance, especially for complex problems where you don't have an answer key, you need a better way to check your work.

You can't just look at the training loss or the residual. You have to check if the price makes sense at the end, if it follows the laws of no-arbitrage, and if its sensitivity numbers are stable. The paper suggests that before we let these AI robots manage our money or hedge our risks, we need to build a "reliability dashboard" that checks all these boxes.

In the end, the paper argues that while AI can extend the reach of mathematical finance into new, complex territories, it requires a new kind of supervision. We need to move from asking "Did the AI follow the rules?" to asking "Is the AI's cake actually edible?" The answer, it turns out, depends on a much broader set of diagnostics than we previously thought.

Technical Summary: Reliability Diagnostics for Physics-Informed Neural Option-Pricing Models

Problem Statement
Physics-Informed Neural Networks (PINNs) have emerged as mesh-free solvers for financial partial differential equations (PDEs), offering the ability to embed governing no-arbitrage PDEs directly into training objectives and compute Greeks via automatic differentiation. However, a critical gap exists between minimizing PDE residual loss and ensuring the reliability of the resulting pricing function. A neural model may achieve low residual loss on sampled collocation points while exhibiting localized pricing errors, unstable sensitivities (Greeks), or violations of economically meaningful shape constraints (e.g., monotonicity and convexity). In quantitative finance, where pricing models are used for hedging and risk management, low residual loss alone is insufficient as a reliability certificate.

Methodology
The paper proposes an a posteriori diagnostic framework to evaluate the reliability of neural option-pricing models. The methodology is grounded in a theoretical "residual–error geometry" perspective and validated through extensive empirical testing.

Residual–Error Geometry: The authors establish a theoretical identity for linear pricing operators, demonstrating that the PDE residual ( $R_\theta$ ) is the differential image of the approximation error ( $e_\theta$ ). Specifically, $R_\theta = \mathcal{L}e_\theta$ . This implies that residuals emphasize high-curvature or locally unstable error components but may fail to detect low-frequency errors. Consequently, small average residuals do not guarantee small pointwise pricing errors.
Diagnostic Framework: The proposed framework aggregates four distinct diagnostic components evaluated on out-of-sample grids:
- Residual Statistics: Mean absolute residual and residual variance to measure PDE inconsistency and spatial heterogeneity.
- Terminal-Condition Consistency: The error between the learned price and the payoff function at maturity ( $\tau=0$ ).
- No-Arbitrage Diagnostics: Measures of monotonicity violations (negative Delta) and convexity violations (negative Gamma) for European calls, which do not require access to the exact solution.
- Greek Instability: The variance of Delta and Gamma in the near-strike region ( $\Omega_{ATM}$ ), targeting stability in the most critical area for hedging.
- Composite Score: A normalized weighted sum of these components serves as a ranking and warning device for model reliability.
Experimental Design:
- Black–Scholes Benchmark: A controlled sweep of 60 PINN configurations varying architecture (depth/width), random seeds, loss weighting ( $\lambda_r, \lambda_T$ ), and input/output scaling. These are compared against a supervised neural network (trained on analytical labels) and a Crank–Nicolson finite-difference reference.
- Heston Pilot Study: A preliminary two-factor stochastic-volatility study to test the framework's extension to operators with mixed derivatives and boundary degeneracy, compared against a semi-analytical Fourier-inversion reference.
- Metrics: Pricing accuracy is measured via Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). Diagnostic efficacy is assessed via Pearson and Spearman correlations with these error metrics.

Key Contributions

Model-Validation Perspective: The paper shifts the focus from architecture design to model validation, emphasizing that residual minimization is necessary but not sufficient for reliable financial pricing.
Theoretical Insight: It formalizes the relationship between PDE residuals and approximation errors, explaining why residual heatmaps and pricing-error heatmaps can display structured, divergent spatial patterns.
Diagnostic Framework: It introduces a composite reliability score that integrates terminal consistency, no-arbitrage structure, and sensitivity stability, moving beyond simple loss monitoring.
Empirical Evidence: Through a sweep of 60 configurations, the study provides data-driven evidence on the limitations of residual-only diagnostics and the benefits of structural constraints.

Results

Supervised vs. PINN: In the Black–Scholes benchmark, a supervised neural network trained on analytical labels achieved significantly lower pricing errors (MAE: 0.19 vs. 2.23) and more stable Greeks than a representative PINN. This indicates that embedding the PDE into the objective does not automatically yield higher accuracy in low-dimensional settings where labels are available.
Diagnostic Correlations: Across the 60 PINN configurations, the proposed full reliability score achieved the strongest Pearson correlation with both MAE (0.8075) and RMSE (0.7779). While mean residuals were informative, they were incomplete; the full score's inclusion of terminal consistency and Greek stability provided a more robust assessment of model quality.
Error Geometry: Heatmaps revealed that PINNs often exhibit broader errors in deep in-the-money and long-maturity regions, whereas supervised models concentrated errors near the non-smooth payoff kink. This supports the theoretical claim that residuals act as differential images, shaping error distribution non-uniformly.
Heston Pilot: The framework successfully applied to the Heston model, where an "improved" configuration (with better scaling and terminal enforcement) reduced pricing errors and the composite score compared to a baseline, demonstrating the framework's applicability to multi-factor, stochastic-volatility operators.

Significance and Claims
The paper claims that the primary contribution is a shift in the validation paradigm for neural pricing models. It argues that in computational regimes where classical methods are expensive (e.g., high-dimensional baskets, XVA, or stochastic volatility), PINNs offer a viable alternative, but their deployment requires rigorous diagnostics beyond loss minimization.

The authors modestly conclude that their work does not claim neural methods replace classical algorithms in low-dimensional settings. Instead, the significance lies in identifying when learned pricing functions can be trusted. The central finding is that reliability in derivative pricing is not merely a matter of fitting prices; it is a structural requirement involving terminal consistency, adherence to no-arbitrage constraints, and the stability of hedging sensitivities. The proposed diagnostic framework serves as a necessary tool to verify these properties before a model is used in risk-management or hedging contexts.