Delta-Hedged Arbitrage Proxy

• Delta-Hedged Arbitrage Proxy is a metric that quantifies deviations from classical no-arbitrage by tracking the realized P&L of a delta-hedged portfolio in imperfect markets.
• It leverages continuous-time hedging, PDE frameworks, and pathwise analysis to detect model mis-specification, discrete rebalancing, and transaction costs across traditional and decentralized finance settings.
• Empirical and algorithmic approaches, such as AMM liquidity PNL calculations and machine learning-based deep hedging, demonstrate its utility in diagnosing arbitrage opportunities under market uncertainty.

A delta-hedged arbitrage proxy quantifies the degree to which a delta-hedged strategy generates persistent positive profit, thereby indicating the presence or magnitude of arbitrage in incomplete or misspecified markets, or under practical frictions. It arises in theoretical settings where the classical no-arbitrage conditions are weakened, such as in models with model uncertainty, strict local martingale dynamics, or structural mis-specification, and also in practical and algorithmic constructs in decentralized finance (DeFi) and machine learning-based hedging. The proxy itself is typically a pathwise, realized or expected accrued profit-and-loss (P&L) of a self-financing strategy deployed in a market that permits some deviation from perfect replication.

1. Pathwise Delta-Hedging and No-Arbitrage Baseline

Conventional financial mathematics establishes that, in continuous-time models with smooth local volatility structure (e.g., Black–Scholes or Föllmer's pathwise calculus), the delta-hedge derived from the unique solution of the parabolic partial differential equation (PDE) or functional Cauchy problem perfectly replicates European and exotic payoffs. This result is rigorous both in the classical probabilistic martingale setting and on a strictly pathwise trajectory level, as formalized in Schied–Voloshchenko’s analysis (Schied et al., 2015).

Let $X_t$ be a $d$-dimensional continuous asset path with quadratic variation determined by a positive-definite volatility matrix $a(t,x) = \sigma\sigma^\top(t,x)$:

• Bachelier: $d\langle X^i, X^j\rangle(t) = a_{ij}(t, X(t)) dt$
• Black–Scholes: $$d\langle X^i, X^j\rangle(t) = a_{ij}(t, X(t)) X^i(t)X^j(t) dt$$

For a European payoff $\Phi(X(T))$, the backward PDE given by

$$\partial_t v + \frac12 \operatorname{tr}[\sigma\sigma^\top \nabla^2_x v] = 0, \quad v(T,x) = \Phi(x)$$

yields delta $\Delta(t) = \nabla_x v(t, X(t))$ so that the self-financing wealth process $v(t, X(t)) = v(0, X(0)) + \int_0^t \Delta(s) dX(s)$ perfectly matches $\Phi(X(T))$.

Under the model class $d$0 and continuous rebalancing, the resulting delta-hedge precludes arbitrage—any apparent arbitrage from delta-hedging in practice is a proxy for deviations from this ideal, such as model mis-specification, discrete rebalancing, or the inclusion of transaction costs, as detailed in (Schied et al., 2015).

2. Delta-Hedged Arbitrage Proxy in Markets with Arbitrage

When the fundamental conditions for no-arbitrage fail (notably when no equivalent local martingale measure exists), delta-hedged trading can generate systematic profits. Ruf (Ruf, 2010) constructs such a proxy in general Markovian models, where the discounted asset price $d$1 evolves as:

$d$2

with a market price of risk $d$3 satisfying $d$4 but possibly with $d$5 a strict local martingale.

The key proxy process here is the pathwise cumulative drift of the delta-hedged portfolio:

$d$6

where $d$7 solves the backward Kolmogorov PDE for the given payoff. This $d$8 grows exactly as the realized delta-hedged P&L in scenarios allowing arbitrage, and dominates zero in bubble models and other settings where $d$9 is not a true martingale (Ruf, 2010).

This framework generalizes to the explicit two-asset delta-hedged arbitrage schemes, such as in (Martynov et al., 2011), where the arbitrage proxy is given by the term:

$a(t,x) = \sigma\sigma^\top(t,x)$0

Here, one constructs a hedged portfolio in two assets sharing a Brownian driver, with instantaneous drift arising whenever the Sharpe ratios differ, resulting in systemic arbitrage.

3. Model Uncertainty and Robust Delta-Hedged Arbitrage Proxy

Under model uncertainty, specifically mean and volatility uncertainty in the style of Peng’s G-calculus, the delta-hedged arbitrage proxy is given by the cumulative finite-variance process $a(t,x) = \sigma\sigma^\top(t,x)$1 in sublinear expectation markets (Xu, 2014). For spot $a(t,x) = \sigma\sigma^\top(t,x)$2 driven by G-Brownian motion with volatility uncertainty $a(t,x) = \sigma\sigma^\top(t,x)$3, super-hedging and sub-hedging prices satisfy nonlinear Black–Scholes–Barenblatt PDEs. In the presence of volatility uncertainty, delta-hedging a short option position yields instantaneous P&L:

$a(t,x) = \sigma\sigma^\top(t,x)$4

where $a(t,x) = \sigma\sigma^\top(t,x)$5 is the Gamma of the option. The process $a(t,x) = \sigma\sigma^\top(t,x)$6 quantifies the maximal cumulative profit/loss from being short a delta-hedged option, serving as a robust, pathwise proxy calibrated using bounds on volatility from data or options-implied surfaces. The ask–bid spread in markets under model uncertainty is expressible as the sum of expected superhedging and subhedging $a(t,x) = \sigma\sigma^\top(t,x)$7-processes, connecting directly to observable market quantities.

4. Algorithmic and Data-Driven Construction in DeFi and Deep Hedging

In decentralized finance, the delta-hedged arbitrage proxy has found concrete application as a risk-adjusted metric in liquidity provision for AMMs. "Liquidity Position PNL" (LPPNL) is defined as the relative mark-to-market gain for LPs after withdrawal, directly offset by a delta-hedged portfolio constructed using vanilla options (Khakhar et al., 2022). For a liquidity position in Uniswap v2, LPPNL as a closed-form function of price movement $a(t,x) = \sigma\sigma^\top(t,x)$8 is

$a(t,x) = \sigma\sigma^\top(t,x)$9

One solves a regularized least-squares problem over a finite set of option payoffs $d\langle X^i, X^j\rangle(t) = a_{ij}(t, X(t)) dt$0 to cancel LPPNL across possible final prices, thereby constructing a delta-hedged strategy whose realized residual P&L tracks arbitrage (fee-capture) cleanly. After hedging, the observable P&L path serves as a transparent proxy for net arbitrage available within the protocol once price risk is neutralized (Khakhar et al., 2022).

A structurally parallel methodology is advanced in "DeFi Arbitrage in Hedged Liquidity Tokens" (Bichuch et al., 2024), where the no-arbitrage price of the CPMM liquidity token is characterized by a perpetual derivative PDE, and the proxy for arbitrage is the observed discrepancy between the token’s AMM mark-to-market price and its replicated value under continuous-time delta-hedging. This framework also supports calibration of an implied volatility through the inverse model value equation, with protocols capable of embedding variable pricing as a direct response to identified arbitrage (Bichuch et al., 2024).

In machine learning-augmented hedging, the discrepancy between standard delta-hedging and deep hedging (where portfolios are learned under generic convex risk measures) can itself be interpreted as a statistical arbitrage proxy (François et al., 2024). Specifically, the P&L of the difference strategy $d\langle X^i, X^j\rangle(t) = a_{ij}(t, X(t)) dt$1, calculated out-of-sample, measures the extent to which deep hedging seeks not merely risk reduction but speculative drift:

$d\langle X^i, X^j\rangle(t) = a_{ij}(t, X(t)) dt$2

is a direct test for statistical arbitrage, with empirical GARCH-market studies showing significant such overlays for lax risk measures (e.g., low-$d\langle X^i, X^j\rangle(t) = a_{ij}(t, X(t)) dt$3 CVaR), but their suppression for loss-only criteria (François et al., 2024).

5. Limitations, Mis-specification, and Real-World Practice

A crucial fact established in (Schied et al., 2015) is that under exact continuous-time hedging within the precise model class (with full information and without frictions), delta-hedged arbitrage cannot arise. All observed or constructed arbitrage proxies, therefore, correspond to:

• Model mis-specification (e.g., misspecified volatility matrix or covariance structure)
• Discrete rebalancing and transaction costs (actual trading can only approximate pathwise continuous trading)
• Market frictions such as slippage, liquidity constraints, or bid-ask spreads

Under ideal conditions, any residual P&L from delta-hedging vanishes. Deviations drive the actionable content of the proxy:

• Persistent positive drift in hedged P&L flags model failure, insufficient risk modeling, or protocol poor design.
• Quantification of proxy values feeds into robust hedging—nonlinear worst-case PDEs (Xu, 2014), explicit pathwise monitoring (Ruf, 2010), or data-driven protocol redesign (Bichuch et al., 2024).

6. Summary Table

Literature Example	Proxy Definition	Context of Arbitrage
(Ruf, 2010) (Ruf)	$d\langle X^i, X^j\rangle(t) = a_{ij}(t, X(t)) dt$4	Non-ELMM Markovian models
(Xu, 2014) (Peng et al.)	$d\langle X^i, X^j\rangle(t) = a_{ij}(t, X(t)) dt$5 process under G-BSDE	Model uncertainty
(Martynov et al., 2011) (Melnikov & Melnikov)	$d\langle X^i, X^j\rangle(t) = a_{ij}(t, X(t)) dt$6	2-asset Sharpe mismatch
(Khakhar et al., 2022, Bichuch et al., 2024) (AMM/DeFi)	Hedged LP P&L	LP token vs. hedge/mark
(François et al., 2024) (François et al.)	$d\langle X^i, X^j\rangle(t) = a_{ij}(t, X(t)) dt$7	Deep hedge vs. delta hedge

Real-world quantification and deployment of delta-hedged arbitrage proxies enables risk managers, protocol designers, and systematic traders to identify, measure, and exploit structural inefficiencies, while providing a rigorous diagnostic of model risk, execution shortfall, and the practical limitations of classical no-arbitrage pricing.