Asymptotically-informed neural networks for Black-Scholes implied volatility computation

Technical Summary: Asymptotically-Informed Neural Networks for Black-Scholes Implied Volatility Computation

Problem Statement
The computation of Black-Scholes implied volatility (IV) is a fundamental task in quantitative finance, essential for option valuation, model calibration, and risk management. While the Black-Scholes pricing formula provides a closed-form solution for option prices given a volatility parameter, the inverse problem—determining the unique volatility $\sigma$ that matches an observed market price—lacks an explicit analytical solution. This inversion becomes a challenging numerical problem, particularly in asymptotic regimes corresponding to extreme option prices, strikes, or maturities (e.g., deep in-the-money/out-of-the-money options or very short/long maturities). In these regions, the inverse map becomes highly sensitive to perturbations in the price; even minor pricing errors can induce large errors in the implied volatility, potentially leading to incorrect hedging decisions or inaccurate model calibration.

Existing numerical methods, such as Newton-Raphson or Brent's algorithm, rely heavily on the quality of the initial guess and may struggle in these extreme regimes. Analytical approximations often lack robustness across realistic parameter ranges, and standard machine learning approaches (generic feed-forward neural networks) often fail to capture the distinct analytical behaviors of the IV function across different volatility regimes, leading to suboptimal accuracy and generalization.

Methodology
The authors propose a new family of neural network architectures that explicitly incorporate the analytical structure of the Black-Scholes inversion problem. The core methodology is inspired by the domain partitioning analysis of Jäckel [30], which identifies that the pricing function exhibits qualitatively different behaviors depending on the volatility regime.

1. Domain Partitioning: The input space, defined by log-moneyness ($A$) and normalized call price ($C$), is partitioned into three distinct regions based on the geometry of the pricing function:

   • Lower Asymptotic Region: Corresponds to very low prices ($C \to 0$).
   • Central Region: Corresponds to the approximately linear behavior near the inflection point of the pricing curve.
   • Upper Asymptotic Region: Corresponds to very high prices ($C \to 1$).

2. Mixture-of-Experts Architecture: The proposed model approximates the implied volatility map $B_{inv}(A, C)$ as a weighted combination of specialized local approximations:
   $$IV(A, C) = f_0(A, C)g_0(A, C) + f_1(A, C)g_1(A, C) + f_2(A, C)g_2(A, C)$$

   • Local Approximators ($g_i$): These are neural networks (or exponential transformations thereof) trained to model the IV behavior within specific regimes.
   • Gating Functions ($f_i$): These are trainable functions that determine the weight of each local approximator. They are designed to satisfy specific asymptotic constraints (e.g., $f_0 \to 1$ as $C \to 0$ and $f_0 \to 0$ as $C \to 1$) to ensure smooth transitions between regimes. The paper investigates various parametric forms for these gates, including polynomial and Gaussian families, with dependencies on both $A$ and $C$ or $C$ alone.

3. Hybrid Refinement: The neural network outputs are used as highly accurate initial guesses for a third-order Householder root-finding scheme. This hybrid approach leverages the speed of the neural network for initialization and the precision of the iterative solver for final refinement.

Key Contributions

1. Regime-Aware Architectures: The design of neural networks that explicitly learn a partition of the price-log-moneyness domain, combining specialized local approximations via gating functions, rather than treating IV inversion as a generic regression problem.
2. Trainable Gating Mechanisms: The introduction of gating functions that automatically learn the transition boundaries between asymptotic and central regimes while respecting known analytical constraints.
3. Superior Accuracy and Generalization: Numerical experiments demonstrate that these architectures consistently outperform standard feed-forward neural networks across a wide range of parameter domains. The improvement is often several orders of magnitude in relative accuracy (MSRE), particularly in challenging asymptotic regimes.
4. Machine-Precision Refinement: The proposed networks provide initial guesses so accurate that a third-order Householder scheme achieves near machine-precision results after only two refinement iterations.
5. Scalability: Unlike analytical approximations that require deriving new formulas for different regimes, the proposed approach scales naturally with model complexity and desired accuracy by simply enlarging the dataset or network size.

Results
The paper presents extensive numerical experiments on three datasets covering different parameter ranges:

• Large-Range Dataset ($A \in [0, 16], B \in [10^{-5}, 7.07]$): The proposed architectures reduced the test Mean Squared Relative Error (MSRE) by approximately four orders of magnitude compared to standard neural networks. Heat maps confirmed that errors were uniformly small across the domain, whereas standard networks exhibited large errors in low-volatility regimes.
• Medium-Range Dataset ($A \in [0, 3], B \in [10^{-7}, 1.22]$): Similar improvements were observed, with the best architectures (specifically those using Gaussian gating functions dependent on both $A$ and $C$) reducing MSRE by over three orders of magnitude.
• Small-Range Dataset ($A \in [0, 10^{-5}], B \in [10^{-5}, 0.18]$): Even in this near one-dimensional regime where standard networks perform reasonably well, the proposed architectures provided systematic improvements, reducing MSRE by a factor of approximately 39.

In all cases, when combined with the Householder refinement, the proposed methods achieved average logarithmic relative errors close to the numerical precision limit (around $-37$ to $-39$), significantly outperforming both standard neural networks and classical root-finding methods like Brent's algorithm, which showed irregular error landscapes in extreme regimes.

Significance
The paper argues that the significance of this work lies in demonstrating that incorporating qualitative analytical knowledge (specifically, the existence of distinct asymptotic and central regimes) into neural network design yields substantially better accuracy and generalization than treating implied-volatility inversion as a generic regression problem. The proposed methodology does not merely replace classical root-finding techniques but acts as a powerful accelerator, placing iterative solvers immediately into their region of rapid convergence. This results in a computationally efficient and highly accurate solver that is particularly robust in the extreme parameter regimes where traditional methods often struggle. The authors suggest that this philosophy of "asymptotically-informed" machine learning could be extended to other calibration problems in quantitative finance involving inverse mappings with multiple regimes and strong non-linearities.