Crisis-Aware Variational Autoencoders for Stress Testing Volatility Surfaces

Technical Summary: Crisis-Aware Variational Autoencoders for Stress Testing Volatility Surfaces

Problem Statement
Stress testing volatility surfaces for exotic derivatives presents a fundamental challenge in risk management. Traditional methodologies suffer from critical limitations when applied to instruments whose values depend on the entire surface structure (e.g., barrier options, variance swaps, TARFs) rather than single points.

• Historical Simulation is bounded by observed market states and cannot generate novel crisis scenarios.
• Parametric (Delta-Gamma) approaches apply uniform shocks (parallel shifts) that destroy the internal consistency of the surface (smile dynamics, term structure) and often fail catastrophically for path-dependent payoffs due to sign-reversal errors in Taylor expansions.
• GARCH-based methods capture volatility clustering but model only a scalar level, failing to represent the complex shape deformations of a surface during stress.
• Expert-designed scenarios provide single point estimates without distributional risk measures (VaR, Expected Shortfall) required for regulatory capital (e.g., FRTB).

The core difficulty lies in generating novel, plausible, and internally consistent stressed surfaces that preserve arbitrage-free conditions (no calendar spread or butterfly arbitrage) while capturing the specific non-linear dynamics of exotic derivatives.

Methodology: The Crisis-Aware VAE
The authors propose a Crisis-Aware Variational Autoencoder (Crisis VAE) designed to generate arbitrage-free volatility surfaces conditioned on explicit crisis severity. The framework is trained on 20,000 synthetic surfaces generated from the Heston stochastic volatility model (with robustness checks on Bates jump-diffusion surfaces), spanning normal, stressed, and crisis regimes.

Key architectural innovations include:

1. Crisis Conditioning via Information Bottleneck: A five-dimensional crisis indicator vector $c$ (encoding VIX-analogue, vol-of-vol, market return, skew, and term structure slope) is passed through a bottleneck layer ($5 \to 2$ dimensions). This compressed vector $\tilde{c}$ is concatenated with the latent code $z$ and fed into the decoder. This design forces a separation of concerns: $\tilde{c}$ controls the severity and regime of the crisis, while the latent code $z$ encodes the specific shape and residual details of the surface.
2. Heavy-Tailed Student-t Prior: Unlike standard VAEs that use a Gaussian prior (which tends to over-smooth and under-represent extremes), the Crisis VAE employs a multivariate Student-t prior with $\nu=4$ degrees of freedom in the latent space. This encourages the generation of fat-tailed scenarios that align with empirical crisis behavior. The prior parameters $\mu(c)$ and $\sigma(c)$ are learned via a separate "PriorNet" conditioned on $c$.
3. Arbitrage-Free Constraints: The training objective includes soft, differentiable penalties for calendar spread violations (non-decreasing total variance) and butterfly spread violations (non-negative risk-neutral density). Additionally, a refined penalty operates directly on call prices (via Black-Scholes inversion) to better approximate the Breeden-Litzenberger condition. Post-generation, surfaces undergo an arbitrage-free SVI (Stochastic Volatility Inspired) fit to ensure strict compliance for continuous instruments like variance swaps.
4. Training Objective: The model maximizes a crisis-conditioned Evidence Lower Bound (ELBO) incorporating reconstruction loss, KL divergence between the Gaussian encoder and the Student-t prior, arbitrage penalties, and a "z-forcing" term to prevent the decoder from ignoring the latent code.

Key Results
The framework was evaluated against four benchmarks (Historical Simulation, Parametric Delta-Gamma, GARCH(1,1) with variance cap, and Expert Regulatory shifts) by pricing a portfolio of exotic instruments (Down-and-In Barrier Put, TARF, Variance Swap) across 500 stress scenarios.

• Precision and Controllability: The Crisis VAE produced the tightest and most controllable stress estimates. For the Down-and-In Barrier Put, the P&L standard deviation was $0.37 (severe regime), significantly lower than Historical Simulation ($1.50) and GARCH ($1.43). The 95% minimum holder gain (VaR) was $6.67, outperforming Historical ($4.92) and GARCH ($4.94).
• Failure of Parametric Approximations: The Delta-Gamma approximation failed structurally for the barrier option, predicting a negative P&L (loss of value) under volatility stress (VaR95 = -$0.38). This "sign reversal" contradicts the theoretical guarantee that higher volatility strictly increases the value of a knock-in put, demonstrating the unsuitability of quadratic approximations for path-dependent exotics.
• Arbitrage Compliance: Generated surfaces satisfied calendar spread constraints with near-zero violations (<0.8%) and butterfly convexity (>97%) on the discrete grid. Post-processing with SVI reduced violations to zero.
• Robustness: The method ranking remained consistent when the training distribution was switched from Heston to Bates jump-diffusion models, indicating the findings are driven by the architecture and instrument properties rather than the specific volatility model.
• Real-Data Validation: An out-of-sample test on a real SPY market surface (February 2024) confirmed the model's ability to reconstruct normal-regime surfaces and map them to the correct crisis coordinates, though the synthetic training data underestimates the steepness of real-world equity skew (a known limitation of the Heston model).

Significance and Claims
The paper claims to establish a principled framework for stress testing that bridges the gap between generative modeling and regulatory requirements. Its primary contributions are:

1. Controllable Stress Generation: The ability to generate distributional risk measures (VaR, ES) at arbitrary confidence levels with explicit control over crisis severity via the indicator vector $c$, satisfying FRTB requirements for forward-looking, plausible scenarios.
2. Surface Consistency: Unlike parallel-shift methods, the Crisis VAE generates surfaces with learned, non-linear deformations (skew steepening, term structure inversion) that are economically consistent and arbitrage-free.
3. Identification of Model Risk: The study provides empirical evidence that parametric delta-gamma approaches are fundamentally flawed for path-dependent exotic derivatives due to sign-reversal failures, advocating for full-surface re-pricing methods.
4. Architectural Novelty: The introduction of a crisis-conditioning bottleneck and heavy-tailed priors specifically tailored for financial stress testing, distinguishing the approach from standard generative models.

Limitations and Scope
The authors explicitly state that the method ranking and numerical results are validated on synthetic data (Heston/Bates). While the framework operates correctly on real market data in normal regimes, the synthetic training distribution does not fully capture real-world crash insurance premia or microstructure effects (e.g., the steep skew observed in real equity options). Consequently, the paper positions the Crisis VAE as a validated methodology on a controlled benchmark and a deployment-ready pipeline pending formal calibration on historical crisis-era option chains (e.g., 2008, 2020), which is identified as a priority for future work. The current results should be viewed as establishing the structural validity of the approach rather than a final, regulator-approved risk model for live markets.