Finance-Informed Neural Network: Learning the Geometry of Option Pricing

The paper introduces Finance-Informed Neural Networks (FINN), a self-supervised machine learning framework that learns option pricing and hedging by minimizing dynamic replication error rather than fitting market prices, thereby ensuring economic consistency, recovering arbitrage-free prices, and enabling robust pricing for assets with no listed options.

The Gist

Imagine you are trying to teach a robot how to price a ticket to a future event, like a concert or a sports game. The ticket's value depends on how popular the event is, how close the date is, and how unpredictable the weather might be.

In the financial world, these "tickets" are called options. For decades, humans have used two main ways to figure out their price:

1. The "Textbook" Way: Using strict math formulas (like the Black-Scholes model). It's like following a recipe. It's reliable, but if the ingredients (market conditions) change in a way the recipe didn't expect, the cake tastes wrong.
2. The "Gut Feeling" Way: Using Machine Learning (AI) to look at past prices and guess the future. It's like a student memorizing the answers to last year's test. It's great at spotting patterns, but it might give you a wrong answer if the test questions change slightly, because it doesn't actually understand the math behind the questions.

The Problem: The "Textbook" way is too rigid, and the "Gut Feeling" way is too risky. It often violates the basic rules of economics (like "you can't make free money"), which makes banks and regulators nervous.

The Solution: This paper introduces FINN (Finance-Informed Neural Network). Think of FINN not as a student memorizing answers, but as a trainee trader learning the job by doing it.

The Core Idea: Learning by "Replicating"

Instead of asking the AI, "What was the price of this option yesterday?" (which is just memorization), FINN asks a different question: "If I hold this option, can I build a portfolio of stocks and cash that perfectly mimics its behavior?"

Here is the analogy:
Imagine you are trying to learn how to balance a broom on your hand.

• Old AI: You watch a video of someone else balancing a broom and try to copy their hand movements. If the broom is slightly different, you fail.
• FINN: You are given a broom and told, "Your goal is to keep this broom from falling. If it falls, you lose points." You don't care about the "price" of the broom; you care about the physics of keeping it upright.

In finance, "keeping the broom upright" means hedging. If you own an option, you buy or sell stocks to cancel out the risk. If your math is wrong, your portfolio loses money. If your math is right, your portfolio grows exactly at the risk-free rate (like a savings account).

FINN's Secret Sauce:
The AI is trained to minimize the "error" in this balancing act. It doesn't look at a list of correct prices. It looks at its own portfolio. If the portfolio drifts away from the safe, risk-free growth, the AI knows it made a mistake and adjusts its internal "brain" (the neural network) to do better next time.

Why This is a Big Deal

1. It Learns the "Rules of the Game" Naturally
Because FINN is trained on the physics of hedging, it naturally discovers the fundamental laws of finance. For example, a rule called Put-Call Parity (which links the prices of two different types of tickets) usually has to be forced into the AI by programmers. With FINN, the AI figures this out on its own because it's the only way to keep the portfolio balanced. It's like a child learning that if you have two apples and give one away, you have one left—not because you were told, but because you tried it and saw the result.

2. It Works When the World is Chaotic
Traditional models assume the market moves smoothly (like a calm lake). But real markets are stormy (volatility changes, crashes happen).

• The Analogy: Imagine driving a car.
  • Old Models: Assume the road is always straight and dry. If it starts raining, the car skids.
  • Old AI: Memorized the route on a sunny day. If it rains, it gets confused.
  • FINN: Learned how to drive by feeling the tires and the steering wheel. It doesn't care if the road is dry or wet; it just adjusts the steering to stay on the road.
  • Result: FINN handles "stormy" markets (like the 2008 crisis or 2020 pandemic) much better than the old methods, keeping the portfolio stable even when the "weather" changes.

3. It Can Price Things That Have No Price Tag
This is perhaps the coolest part. Usually, to price an option, you need a market where people are already trading that option. But what if you have a brand-new asset (like a new tech stock or a weird ETF) that has no options market yet?

• Old Way: You can't price it. You have to guess.
• FINN: You just feed it the history of the stock's price movements. FINN learns the "personality" of that stock (how it jumps, how it drifts) and instantly generates a fair price and risk profile for options on that stock, even though no one has ever traded them before. It's like a chef tasting a new ingredient and instantly knowing how to cook a perfect dish with it, without ever having seen a recipe.

The Bottom Line

FINN changes the game by shifting the focus from "What is the price?" to "How does the price behave?"

It combines the best of both worlds:

• The rigor of financial theory (it can't break the laws of economics).
• The flexibility of modern AI (it can adapt to messy, real-world data).

It's like giving a robot a deep understanding of physics and then letting it learn to drive a car in any weather, on any road, without ever needing a human to tell it the speed limit. It's a smarter, safer, and more adaptable way to manage financial risk.

Technical Summary

Here is a detailed technical summary of the paper "Finance-Informed Neural Network: Learning the Geometry of Option Pricing."

1. Problem Statement

The paper addresses a fundamental dilemma in quantitative finance: the trade-off between theoretical consistency and empirical performance in option pricing and hedging.

• Traditional Models (e.g., Black–Scholes, Heston): These are grounded in economic theory (no-arbitrage, dynamic replication) and are interpretable. However, they often rely on restrictive parametric assumptions (e.g., constant volatility) that fail to capture complex market behaviors, leading to poor empirical accuracy in stochastic volatility environments.
• Standard Machine Learning (ML) Models: These can fit market data with high accuracy but often operate as "black boxes." They typically minimize statistical prediction errors (fitting observed prices) rather than economic criteria, which can result in prices and sensitivities (Greeks) that violate fundamental no-arbitrage principles (e.g., put-call parity).

The core challenge is to develop a framework that learns from data while strictly adhering to the economic laws of replication and no-arbitrage, without requiring labeled option price data or specific parametric assumptions about asset dynamics.

2. Methodology: The FINN Framework

The authors propose the Finance-Informed Neural Network (FINN), a hybrid framework that embeds financial theory directly into the neural network's learning objective.

Core Concept: Self-Supervised Replication

Instead of training on observed option prices (supervised learning), FINN uses a self-supervised replication objective. The network is trained to minimize the hedging (replication) error of a dynamically managed portfolio.

• The Logic: In a complete, arbitrage-free market, a self-financing portfolio hedged against the underlying asset (and potentially a second derivative) must earn the risk-free rate.
• The Loss Function: The network $g_\theta$ is trained to minimize the mean squared deviation between the realized portfolio value change and the risk-free growth:
  $$\mathcal{L}(\theta) = \mathbb{E}\left[ \sum_{n=0}^{N-1} \left( \Pi_{t_{n+1}} - e^{\int_{t_n}^{t_{n+1}} r_s ds} \Pi_{t_n} \right)^2 \right]$$
  Where $\Pi_t$ is the value of a portfolio shorting the option and holding hedging instruments (underlying asset $S_t$ and potentially a second option $H_t$).

Key Technical Features

1. Automatic Differentiation for Greeks: The network architecture uses smooth activation functions (e.g., tanh, swish) to ensure that first-order (Delta, $\Delta$) and second-order (Gamma, $\Gamma$) sensitivities can be computed via automatic differentiation. These sensitivities are used to construct the hedge ratios endogenously.
2. Delta-Gamma Hedging: The framework extends beyond simple delta hedging. By incorporating a second liquid derivative, FINN can enforce Gamma neutrality, allowing for more robust risk management.
3. American Options: For American options, the framework is extended to handle the early-exercise feature. The loss function includes a penalty for violating the constraint $g(t, S) \geq f(S)$ (where $f$ is the intrinsic payoff) and restricts the hedging loss to the continuation region.
4. No Parametric Assumptions: FINN does not assume a specific stochastic differential equation (SDE) for the underlying asset. It learns the pricing operator directly from price transitions, making it applicable even when no options market exists.

Theoretical Guarantees

The paper provides theoretical proofs showing that:

• Minimizing the replication residual in the continuous-time limit is mathematically equivalent to solving the standard no-arbitrage pricing PDE (e.g., Black–Scholes).
• Under standard regularity conditions, the learned function converges to the unique arbitrage-free price and its sensitivities.

3. Key Contributions

1. A New Learning Paradigm: FINN shifts the focus from "fitting prices" to "learning a pricing operator" based on replication and risk-control principles.
2. Endogenous No-Arbitrage: Fundamental relationships like Put-Call Parity emerge naturally from the training process without being explicitly imposed as constraints.
3. Robustness to Model Misspecification: FINN performs accurately in stochastic volatility environments (Heston model) where classical models struggle, and it generalizes well to out-of-sample data.
4. Applicability to Illiquid Markets: The framework can be trained directly on historical spot prices to generate coherent option prices and Greeks for assets with no listed options (e.g., new ETFs or niche assets), solving a major practical problem in valuation.

4. Empirical Results

The authors evaluated FINN across several scenarios:

• Black–Scholes (GBM) Environment: FINN accurately recovers analytical Black–Scholes prices and Greeks, with relative mean absolute deviations (RMAD) typically below 4%.
• Stochastic Volatility (Heston): In environments with time-varying volatility, FINN outperforms standard parametric approaches, maintaining low pricing errors and stable hedge ratios without needing closed-form solutions.
• Put-Call Parity: The model naturally satisfies put-call parity across different volatility regimes, demonstrating that it has internalized the economic structure of arbitrage-free pricing.
• Delta-Gamma Hedging: FINN-based delta-gamma hedging significantly reduces hedging errors compared to delta-only strategies.
• Stress Tests (Transaction Costs & Volatility Shocks): In stress tests using real market data (2008, 2011, 2020) with transaction costs, FINN-based hedging strategies demonstrated lower downside risk (VaR/CVaR) and lower transaction costs compared to traditional Black–Scholes delta hedging. FINN produces smoother hedge adjustments, avoiding the aggressive rebalancing that leads to high costs in volatile markets.
• Implied Volatility (IV) Surface Reconstruction: When reconstructing IV surfaces, FINN outperformed a hybrid Heston calibration. The Heston model exhibited a systematic negative bias (underestimating volatility) due to structural misspecification, whereas FINN captured the market geometry more accurately, showing superior out-of-sample adaptability.
• Assets Without Options: FINN was successfully trained on historical spot prices of three newly issued instruments (TACO, GTEN, COPL) to generate coherent call/put price and delta surfaces, demonstrating its utility for pricing bespoke derivatives where no market data exists.

5. Significance and Conclusion

The paper presents a significant advancement in financial machine learning by bridging the gap between theoretical finance and data-driven methods.

• Practical Impact: FINN offers a scalable tool for pricing, hedging, and risk management in both liquid and illiquid markets. It is particularly valuable for institutions dealing with bespoke derivatives or assets lacking active option markets.
• Theoretical Shift: By reframing option pricing as the learning of a pricing operator via replication consistency, FINN ensures that AI-driven financial models are not just statistically accurate but also economically sound and interpretable.
• Future Directions: The framework opens avenues for multi-asset pricing, incorporating jumps and stochastic interest rates, and integrating with online learning for dynamic market adaptation.

In summary, FINN provides a robust, theory-informed alternative to traditional parametric models and black-box ML, delivering accurate pricing and superior risk management performance across diverse market regimes.