Adaptive Shrinkage Calibration of Fractional Order and Volatility in the Time-Fractional Black–Scholes Model: Evaluation under Symmetric and Asymmetric Losses

Technical Summary: Adaptive Shrinkage Calibration in the Time-Fractional Black–Scholes Model

Problem Statement
The paper addresses the inverse calibration problem for the time-fractional Black–Scholes (TFBS) model, specifically the joint estimation of the fractional order ($\alpha$) and volatility ($\sigma$) from option prices. While fractional calculus offers a non-local framework to encode memory effects in price dynamics via the Caputo derivative, the practical application faces three critical challenges:

1. Dimensional Inconsistency: Standard formulations often lack explicit time scaling, making the fractional order dependent on the unit of time.
2. Numerical Instability: The inverse problem is ill-conditioned, leading to high variance in raw estimates, particularly when $\alpha$ and $\sigma$ exhibit strong local dependence (compensation effects in the price surface).
3. Boundary and Verification Issues: Many numerical schemes fail to verify the fractional memory convolution or handle far-field boundaries correctly in the fractional regime, often relying solely on verification at the classical limit ($\alpha=1$).

The authors aim to determine whether noisy, maturity-specific inverse estimates can be stabilized using shrinkage techniques and whether such stabilization translates to improved out-of-sample pricing performance under both symmetric and asymmetric loss functions.

Methodology

• Model Formulation:
  The authors propose a dimensionally balanced TFBS model. To ensure dimensional coherence, a scaling factor $\rho_\alpha = \tau_\star^{\alpha-1}$ (where $\tau_\star = 1$ year) is introduced, ensuring the fractional derivative term has units of price per year. The model incorporates a "fractional far-field" boundary condition derived from the associated fractional reaction equation, resulting in a Mittag–Leffler discount term rather than the classical exponential decay.

• Numerical Scheme:
  The forward problem is solved using an explicit $L1$ approximation for the Caputo time derivative combined with Crank–Nicolson spatial averaging. The scheme is verified in two regimes:

  1. Classical Limit ($\alpha=1$): Comparison against the closed-form Black–Scholes solution.
  2. Fractional Eigenmode: Verification against an exact solution involving the Mittag–Leffler function ($E_\alpha$) to confirm the correct implementation of the memory convolution and to observe the reduced temporal convergence rate (often near first-order) due to weak initial singularities on uniform meshes.

• Inverse Calibration and Diagnostics:
  Raw estimates ($\hat{\theta}_j$) for each maturity $j$ are obtained by minimizing a standardized training loss on a precomputed price grid. The authors compute local calibration uncertainty matrices ($\hat{\Sigma}_j$) using the Jacobian of the price surface. Key diagnostics include the condition number of the Hessian and the correlation ($\rho_j$) between $\alpha$ and $\sigma$, which often indicates strong local dependence.

• Shrinkage Estimators:
  To stabilize the raw estimates, four methods are compared:

  1. Fixed Shrinkage Estimator (FSE): Shrinks estimates toward a global componentwise median (batch rule).
  2. Causal Dynamic-Target Shrinkage (DTSE): Shrinks toward an updated causal target using only past maturities.
  3. Full-Matrix Adaptive Shrinkage Estimator (ADSE): A state-space prediction-correction approach using full covariance matrices to account for the correlation between $\alpha$ and $\sigma$.
  4. Validation-Tuned Penalized Smoother (VPS): A batch smoother penalizing deviations between adjacent maturities, weighted by inverse marginal variances.

  Hyperparameters for all shrinkage methods are selected via validation-only tuning (minimizing loss on a held-out validation set), ensuring the test set remains strictly unseen.

• Evaluation Framework:
  The study employs a Monte Carlo design with ten maturities, varying strike counts, and noise levels. It evaluates performance under:

  • Parameter Loss: Standardized squared error and General Entropy (GE) loss (for $q=1$ and $q=-1$) to assess bias-variance trade-offs and asymmetric penalties.
  • Price Loss: Held-out standardized option price error and positive-price entropy loss.
  • Stress Paths: Constant, smooth, jump, and oscillatory parameter paths to test robustness against excessive smoothing.

Key Results

• Monte Carlo Simulations:
  The simulations demonstrate that shrinkage estimators (particularly ADSE and VPS) can significantly reduce parameter risk (variance) compared to raw calibration when the true parameter paths are smooth or constant. However, under "stress paths" involving sharp jumps or oscillations, excessive smoothing can increase bias, indicating that no single estimator dominates all scenarios. The full-matrix ADSE generally outperforms diagonal-gain approaches when $\alpha$ and $\sigma$ are highly correlated.

• Empirical Application (S&P 500, May 11, 2012):
  Applying the framework to cleaned S&P 500 call option data yields specific findings:

  • Boundary Behavior: All raw fractional-order estimates ($\hat{\alpha}_j$) converged to the classical boundary ($\alpha=1$).
  • Identification: The joint calibration was found to be locally ill-conditioned, with strong correlation between $\alpha$ and $\sigma$.
  • Stabilization: While the ADSE could mathematically push estimates below 1 via regularization, the data provided no evidence that $\alpha < 1$ was required for this specific cross-section.
  • Performance: The gains in held-out price loss were described as "modest," suggesting that while stabilization reduces parameter volatility, it does not drastically improve pricing accuracy for this specific dataset.

Significance and Claims

The paper modestly claims to provide a transparent, reproducible framework for stabilizing fractional order calibration rather than proving the necessity of fractional dynamics for the specific dataset used.

1. Methodological Rigor: The work establishes a dimensionally consistent model with a fractional far-field boundary and a numerically verified $L1$–Crank–Nicolson scheme that explicitly handles the weak initial singularities of fractional diffusion.
2. Diagnostic Utility: It demonstrates that raw inverse estimates can be highly unstable and that shrinkage is a necessary diagnostic and regularization tool. The full-matrix adaptive approach is highlighted as superior for handling the specific correlation structure between fractional order and volatility.
3. Empirical Caution: The study concludes that for the specific S&P 500 cross-section analyzed, the data does not support a departure from the classical Black–Scholes limit ($\alpha=1$). The authors argue that the result may reflect model misspecification, the volatility smile, or local confounding, rather than the absence of memory effects in the market.
4. Validation Strategy: The paper emphasizes the importance of validation-controlled tuning to prevent overfitting and the distinction between minimizing parameter risk and maximizing pricing performance.

In summary, the paper contributes a robust numerical and statistical infrastructure for fractional option pricing, validating that while shrinkage stabilizes parameter estimates, the empirical evidence for fractional dynamics in standard equity options remains inconclusive without further data or model extensions.