Differential Machine Learning for 0DTE Options with Stochastic Volatility and Jumps

This paper introduces a differential machine learning framework for pricing 0DTE options under stochastic volatility and jumps, which utilizes a specialized network architecture and a three-stage training procedure to simultaneously compute accurate prices and Greeks while significantly outperforming traditional Fourier-based benchmarks in speed and stability.

The Gist

Imagine you are a trader standing on a busy street corner, trying to sell tickets to a show that starts right now. These are "0DTE" options: financial contracts that expire in the same day they are bought. Because they expire so quickly, their value is incredibly sensitive to tiny movements in the stock price and sudden, unexpected shocks (like a celebrity scandal or a surprise news headline).

The paper by Takayuki Sakuma is about building a super-smart, instant calculator to price these tickets and figure out how to protect yourself if the market goes crazy.

Here is the breakdown using simple analogies:

1. The Problem: The "Fragile Glass" and the "Surprise Rock"

In the world of finance, predicting the price of these ultra-short-term tickets is like trying to balance a glass of water on a wobbly table while someone throws rocks at it.

• The Wobble (Volatility): The stock price jumps around wildly.
• The Rocks (Jumps): Sometimes, the price doesn't just drift; it teleports up or down due to sudden news.
• The Challenge: Traditional math is too slow to calculate the price every time a rock hits. If you can't calculate it fast enough, you lose money. Also, near the "money" (where the ticket price equals the stock price), the math gets so unstable it's like trying to measure the thickness of a hair with a ruler.

2. The Solution: A "Twin-Brain" Neural Network

The author uses a technique called Differential Machine Learning. Think of this not as a single calculator, but as a twin-brain system trained to do two things at once:

1. Brain A (The Price): Guesses the ticket price.
2. Brain B (The Sensitivity): Guesses how much the price will change if the stock moves a tiny bit (these are called "Greeks," like Delta and Gamma).

Usually, you have to calculate the price first, then do a second, messy calculation to find the sensitivity. This new method does both in one single step, like a chef who can taste a soup and instantly know exactly how much salt to add, without needing a second spoon.

3. The Secret Sauce: Three Tricks to Make it Work

Trick A: The "Magic Lens" (Variance Correction)

Instead of asking the AI to learn the price from scratch (which is hard because the price behaves wildly near expiration), the AI is given a standard Black-Scholes formula (a basic, well-known recipe) as a starting point.

• The Analogy: Imagine you are trying to draw a perfect circle. Instead of starting with a blank page, you are given a slightly imperfect circle and asked to "fix" the wobbles.
• The AI only learns the correction needed to make the basic recipe perfect. As the expiration time gets closer to zero, the AI is forced to stop correcting and just hand over the final result (the payoff). This keeps the math stable.

Trick B: The "Two-Stage Detective" (Identifying the Jumps)

This is the most clever part. In the math, the "smooth drift" of the stock and the "sudden jumps" can look very similar to a computer. If you just tell the computer, "Make the final equation balance," it might cheat: it could say, "Oh, the stock moved smoothly," when actually it was a jump, just to make the numbers add up.

• The Fix: The author introduces a second network specifically to hunt for the "Jump" part.
• The Three-Stage Training:
  1. Stage 1: Teach the Price Brain to get the basics right.
  2. Stage 2: Freeze the Price Brain and teach the Jump Brain to recognize what a "jump" actually looks like (using a reference calculator).
  3. Stage 3: Let them work together, but keep an eye on them to make sure they aren't cheating each other.

Trick C: The "Safety Net" (No-Arbitrage Rules)

The AI is also taught the basic laws of the financial universe: "You can't make free money," "Prices can't be negative," and "If the stock goes up, the option price must go up." If the AI breaks these rules, it gets a "penalty" during training. This ensures the AI doesn't learn weird, impossible patterns.

4. The Results: Fast, Accurate, and Safe

The paper tested this system against a traditional, very slow, high-precision math method (Fourier transform).

• Speed: The AI is 30 to 47 times faster than the traditional method. It's the difference between waiting for a slow train and taking a helicopter.
• Accuracy: It predicts the price just as well as the slow method, but it predicts the sensitivity (how much you should buy or sell to stay safe) much better.
• Hedging: When they simulated a "stress test" (a day with huge market jumps), the AI's strategy kept the trader's profit/loss stable. It didn't panic when the "rocks" were thrown.

Summary

The paper presents a new way to train an AI to be a super-fast, ultra-precise financial navigator for the most volatile, short-term options. By teaching the AI to learn corrections to a standard recipe and by using a special "detective" network to spot sudden market jumps, they created a tool that is fast enough for real-time trading but smart enough to handle the chaos of the market.

In one sentence: They built a "smart twin-brain" calculator that instantly prices risky, expiring-today options and tells you exactly how to protect yourself, even when the market suddenly jumps.

Technical Summary

Here is a detailed technical summary of the paper "Differential Machine Learning for 0DTE Options with Stochastic Volatility and Jumps" by Takayuki Sakuma.

1. Problem Statement

The paper addresses the computational challenges associated with pricing and hedging Zero-Days-to-Expiry (0DTE) options under a Stochastic Volatility Jump-Diffusion (SVJD) model (specifically the Bates model).

• Market Context: 0DTE options dominate trading volume but exhibit extreme sensitivity to intraday jumps and large Gamma exposures near the At-The-Money (ATM) region.
• Numerical Challenges:
  • Instability: Near maturity ($\tau \to 0$) and near the money, option Greeks (especially Gamma) become numerically unstable and singular.
  • Identifiability: In Partial Integro-Differential Equation (PIDE) solvers using residual penalties, the jump term is often unidentifiable. The optimizer can minimize the total residual by allowing the jump network to "cancel out" errors in the diffusion term without actually learning the correct jump dynamics.
  • Speed: Traditional Fourier-based methods are accurate but computationally expensive for real-time, high-frequency hedging requiring repeated evaluations of prices and Greeks.

2. Methodology

The authors propose a Differential Machine Learning (DML) framework that computes prices and Greeks in a single network evaluation. The architecture and training strategy involve three core innovations:

A. Architecture: Variance-Gated Black-Scholes Formulation

Instead of predicting option prices directly, the network learns a variance correction applied to the Black-Scholes formula.

• Input: Log-moneyness ($x$), time-to-maturity ($\tau$), variance ($V$), and model parameters ($\kappa, \theta, \sigma_v, \rho, \lambda, \mu_J, \sigma_J$).
• Effective Variance: The network outputs a correction $\Delta V_\phi(x)$, scaled by a maturity-gating function $g(\tau) = 1 - \exp(-\tau/\tau_0)$.
  $$V_{eff} = \max\{V + g(\tau)\Delta V_\phi(x), \epsilon\}$$
• Output: The price is calculated as $C_{BS}(S, K, \tau, \sqrt{V_{eff}})$.
• Benefit: As $\tau \to 0$, $g(\tau) \to 0$, forcing the correction to vanish. This ensures the network naturally satisfies the payoff boundary condition ($\max(S-K, 0)$) and reduces the approximation difficulty in the near-singular regime.
• Greeks: Calculated via Automatic Differentiation (AD) of the network output with respect to inputs.

B. Twin-Network Design

To handle the non-local jump operator, the system uses two coupled networks:

1. Price Network ($\phi$): Outputs the variance correction and price.
2. Jump-Operator Network ($\psi$): Explicitly approximates the compensated jump operator $J_\psi(x)$.

C. Three-Stage Training Procedure

To solve the identifiability issue where the jump network might simply cancel diffusion errors, the authors introduce a three-stage training schedule:

1. Stage 1 (Price & Greeks): Train the price network ($\phi$) using supervised price and Greek targets (Delta, Gamma, Vega) plus no-arbitrage penalties. The jump network ($\psi$) is frozen.
2. Stage 2 (Jump Reference): Freeze the price network. Train the jump network ($\psi$) to match a numerical proxy of the compensated jump term ($J_{ref}$) computed from a high-precision Fourier pricer. This forces the jump network to learn the true jump dynamics rather than just minimizing residuals.
3. Stage 3 (Joint Refinement): Jointly train both networks. The loss function includes:
   • Price and Greek supervision.
   • PIDE residual penalty ($L_{diff} + \lambda J_\psi$).
   • Self-Consistency Penalty: A term penalizing the mismatch between the learned jump network output and a low-order numerical evaluation of the jump operator applied to the current price network.

3. Key Contributions

• Identifiable Jump Learning: The paper demonstrates that standard residual-penalty methods fail to learn the correct jump operator in SVJD models because the jump term can absorb diffusion errors. The proposed three-stage training with explicit jump supervision resolves this, ensuring the learned jump term matches the model-implied integral.
• Maturity-Gated Variance Correction: By embedding the Black-Scholes structure and gating the variance correction, the method stabilizes training for ultra-short maturities where standard neural PDE solvers often fail.
• Single-Pass Efficiency: The model outputs prices and all major Greeks (Delta, Gamma, Vega) simultaneously via AD, eliminating the need for finite difference approximations or multiple pricing loops.
• Robust Loss Functions: The use of Huber loss for Gamma and jump terms, combined with no-arbitrage constraints (convexity, monotonicity), mitigates numerical noise in deep Out-of-The-Money (OTM) regions.

4. Numerical Results

Experiments were conducted using the Bates SVJD model with simulated data.

• Accuracy:
  • Prices: The three-stage model (Model D) achieved Price RMSE comparable to one-stage baselines (Models B and C) and slightly better than price-only models.
  • Greeks: The DML approach significantly improved Delta and Vega accuracy compared to price-only training. While Gamma remained the most challenging target, the three-stage model showed the best tail-error quantiles for Gamma.
  • Jump Term: Crucially, Model D (three-stage) achieved a 10x lower error in the learned jump operator ($J_\psi$) compared to Model C (residual-only), proving that the jump term was correctly identified rather than just acting as a residual canceller.
• Speed:
  • The DML pricer was substantially faster than the Fourier-based benchmark.
  • For a batch of $50,000$ options, the DML model was 47.4x faster than the Fourier method while maintaining similar price accuracy.
• Hedging Performance:
  • In one-day Delta-hedging simulations under stressed parameters (high jump intensity), the P&L distribution of hedges using DML Greeks was nearly identical to those using "true" Fourier Greeks.
  • The method proved robust even when realized jumps occurred, though jumps naturally thickened the P&L tails.

5. Significance

This work provides a practical solution for the rapid pricing and hedging of 0DTE options, a critical asset class in modern markets.

• Operational Efficiency: The massive speedup over Fourier methods enables real-time risk management and high-frequency rebalancing strategies that were previously computationally prohibitive.
• Theoretical Insight: It highlights a critical flaw in standard "physics-informed" neural network approaches for jump-diffusion models (the non-identifiability of the jump term) and offers a concrete architectural and training solution.
• Market Applicability: The method maintains stability in the most difficult regime (ultra-short maturity, ATM, with jumps), making it suitable for institutional trading desks requiring accurate, instantaneous Greeks for hedging.

Conclusion: The paper successfully bridges the gap between high-fidelity stochastic modeling (SVJD) and the computational speed requirements of 0DTE trading, demonstrating that with careful architectural design and staged training, machine learning can outperform traditional numerical methods in both speed and stability without sacrificing accuracy.