Observation-Anchored Arbitrage Repair for Sparse Implied Volatility Surface Completion
Abstract
Completing dense implied volatility surfaces (IVS) from sparse option quotes is central to derivatives pricing and risk management. A financially admissible surface must satisfy calendar and butterfly no-arbitrage conditions, yet existing completion methods do not enforce these constraints while preserving observed market quotes. We propose a deterministic, observation-anchored two-step repair algorithm enforcing calendar constraints in total-variance space and butterfly constraints in normalized call-price space, modifying only unobserved grid cells. The algorithm is evaluated across five completion backends—linear interpolation, cubic spline, SABR, U-Net, and a conditional diffusion model—using real S&P 500 index (SPX) option data from OptionMetrics Ivy DB over 252 out-of-sample trading days at three holdout levels. Calendar violation rates were reduced to near-zero across all backends and sparsity levels. Butterfly violations were partially reduced, bounded below by an empirically observed floor of approximately 5.4% attributable to finite-precision rounding in the IV-to-call-price round-trip. Cubic spline with repair achieved the lowest reconstruction error at all holdout levels; the diffusion backend showed increased reconstruction error after repair alongside substantial calendar violation reduction. Ablation analysis confirmed that calendar and butterfly repair interact asymmetrically and that the combined pipeline is necessary. These findings support observation-anchored post-processing as a practical, model-agnostic approach for enforcing financial consistency in sparse IVS completion.