Static versus Dynamic Arbitrage Bounds on Multivariate Option Prices

Alexandre d'AspremontarXiv 2004 · DOI

Abstract

In single-asset option pricing, the inability of the Black-Scholes model to fit a market price typically signals a static arbitrage opportunity. This paper demonstrates that this intuition fails in the multivariate setting. For basket options, there exists a significant "gray area" where market prices are mathematically arbitrage-free yet cannot be represented within a multivariate Black-Scholes framework. By employing Semidefinite Programming (SDP) and Linear Programming (LP) relaxations, this research quantifies the gap between model-specific bounds and general static arbitrage bounds, highlighting the inherent model risk in multivariate derivatives trading.

Introduction

In the classical Black-Scholes framework for a single asset, the relationship between model calibration and market reality is tightly coupled. If one cannot find an implied volatility that matches a market price, it is a definitive sign of a "free lunch" or static arbitrage—a guaranteed profit through buying and holding.

However, modern financial markets increasingly rely on multivariate instruments, such as basket options (options on an index or a collection of assets) and spread options. In these settings, the relationship between models and arbitrage becomes significantly more complex. Calibrating a multivariate Black-Scholes model is computationally intensive, and verifying the absolute absence of arbitrage in a multivariate distribution is known to be an NP-Hard problem. This paper investigates the discrepancy between the range of prices a model can produce and the broader range of prices that are simply "legal" under the laws of no-arbitrage.

Methodology

The research utilizes three distinct mathematical frameworks to establish price boundaries:

1. Model Bounds via Semidefinite Programming (SDP)

To determine the range of prices possible within a multivariate Black-Scholes (lognormal) framework, the author formulates the calibration problem as a Semidefinite Program. This allows for the efficient calculation of the upper and lower price limits that a Black-Scholes-style model can reach for a given basket, assuming the underlying assets follow a joint lognormal distribution.

2. Static Arbitrage Bounds via LP Relaxation

Because determining exact arbitrage-free bounds is NP-Hard, the author employs a Linear Programming relaxation. By discretizing the necessary conditions for a valid price—specifically positivity, monotonicity, and joint convexity—into a finite linear program, the study provides a computationally tractable estimate of the total arbitrage-free price range.

3. The Moment Approach

For higher precision, the author introduces a method based on moment theory and payoff semigroups. This approach treats option prices as moments of an underlying probability distribution. This leads to a hierarchy of increasingly precise SDP relaxations that converge toward the true arbitrage-free bounds.

Experimental Design

Due to the complexities of high-dimensional market data, the study utilizes controlled simulations:

  • Discrete 2D Model: A simplified six-point distribution used to validate the tightness and accuracy of the LP relaxations.
  • 5-Asset Multivariate Lognormal Model: A more complex simulation used to generate "arbitrage-free" prices. This model serves as the primary testbed for comparing the restrictive nature of the Black-Scholes model against the theoretical limits of static arbitrage.

Key Insights

The Multivariate "Gray Area"

The most significant finding is the quantification of the "gray area": a vast set of prices that do not allow for static arbitrage but are impossible to replicate using a multivariate Black-Scholes model. This suggests that the Black-Scholes framework is surprisingly restrictive when applied to baskets.

Challenging Market Intuition

In single-asset pricing, at-the-money (ATM) calls are generally robust and less sensitive to specific model assumptions. The author proves this intuition is incorrect for multivariate assets. Even for ATM basket calls, there is a substantial gap between the prices allowed by the model and those allowed by the absence of arbitrage.

Information Efficiency

The numerical results indicate that the "legal" price range for a target option narrows significantly with the addition of only a few basket prices. This suggests that market participants do not require an infinite data set to establish meaningful risk boundaries, provided they use the correct optimization techniques.

Conclusion and Implications

The research highlights a high degree of "model risk" for practitioners. If a trader relies exclusively on a multivariate Black-Scholes model, they may incorrectly identify a market price as an arbitrage opportunity when, in reality, the model is simply too rigid to capture the underlying distribution.

While the study demonstrates that LP and SDP relaxations are computationally feasible for risk management, it also identifies limitations in current solvers (such as SeDuMi) when handling complex moment-based relaxations. Future work is required to scale these techniques to higher-dimensional baskets, such as the S&P 500, and to incorporate more complex models that account for volatility smiles and skewness.

Summary updated

Abstract

We compare static arbitrage price bounds on basket calls, i.e. bounds that only involve buy-and-hold trading strategies, with the price range obtained within a multi-variate generalization of the Black-Scholes model. While there is no gap between these two sets of prices in the univariate case, we observe here that contrary to our intuition about model risk for at-the-money calls, there is a somewhat large gap between model prices and static arbitrage prices, hence a similarly large set of prices on which a multivariate Black-Scholes model cannot be calibrated but where no conclusion can be drawn on the presence or not of a static arbitrage opportunity.
BibTeX
@misc{daspremont2004static,
  author = {Alexandre d'Aspremont},
  title = {{Static versus Dynamic Arbitrage Bounds on Multivariate Option Prices}},
  year = {2004},
  eprint = {cs/0407029},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/cs/0407029},
}
Figure