A Unified Risk-Averse Framework for XVAs and Hedging Costs

Abstract

We develop a unified stochastic-control framework for XVA management in which a derivatives desk jointly optimises its hedge portfolio, funding strategy, and collateral policy subject to counterparty credit risk, book-funding costs, and market-impact frictions. Following ?, the desk cash dynamics include an explicit book-funding cost −f_t v_t dt — the direct source of FVA — together with the standard received-collateral convention C_t > 0. Under exponential (CARA) utility the HJB decomposes into a clean price w_0 and a nonlinear XVA adjustment u. A perturbative expansion in risk-aversion and friction strength recovers the complete classical XVA stack — CVA, FVA, ColVA — at leading order with discount rate f + λ^P ζ_C and real-world P-expectations, while HVA and variance penalties appear at first order. A quadratic reduction yields an exact Riccati system with a −α(t)M(t) coupling term absent from standard LQR theory; this encodes the interaction between book-funding cost and default linearisation. Under constant coefficients the system is solved in closed form: optimal delta and credit-hedge rates are explicit mean-reversions whose targets, speeds, and steady-state values are given in terms of model primitives.

Summary

This paper formulates the XVA management problem as a stochastic control problem under the real-world measure P with CARA utility. The key structural choices are: (i) including a book-funding cost term $-f_t{v}_{t}dt$ in the desk wealth SDE, (ii) working under P rather than Q, and (iii) using exponential utility to obtain a PDE (HJB) for the total derivative MTM $v$, which decomposes as clean price $w_0$ plus XVA adjustment $u$. The HJB for $u$ (eq. 8) is a semilinear PDE with quadratic gradient terms, an exponential default jump, and the book-funding discount $-f_{t}u$.

A perturbative expansion ($\gamma =\epsilon \rho$, small risk-aversion and frictions) recovers the classical XVA stack at leading order under P-world quantities: CVA, FVA, and ColVA appear at $O(1)$ with discount factor $D^P$ at rate $f+{\lambda }^P{\zeta }_C$, while HVA (hedging valuation adjustment from market-impact frictions) and diffusive variance penalties appear at $O(\epsilon )$. Under six reduction hypotheses (pure diffusion, no collateral control, frozen coefficients, quadratic frictions, affine delta, quadratic default surrogate), the HJB reduces to a quadratic PDE solved exactly by a Riccati system. The resulting closed-form hedging rates are mean-reversions with tanh-shaped gains toward time-dependent targets.

Key Contributions

• HJB equation for total derivative value $v$ under real-world measure P (eq. 5), and for the XVA adjustment $u$ (eq. 8)
• Well-posedness of the associated Qexp BSDE (Theorem 1) and viscosity characterisation (Theorem 2)
• Perturbative recovery of the full XVA stack (CVA^P, FVA^P, ColVA^P) at leading order with correct P-world discount $f+{\lambda }^P{\zeta }_C$ (eq. 13)
• Identification that HVA enters at $O(\epsilon )$, not $O(1)$, under the $\gamma =\epsilon \rho$ scaling
• Matrix Riccati ODE (eq. 18) with novel $-\alpha (t)M(t)$ coupling encoding book-funding/default interaction
• Closed-form scalar Riccati solutions with tanh profiles (eq. 25-26) and explicit hedge targets (eq. 29-31)
• XVA overlay formulation (Section 7) for practical desk implementation

Methodology

The paper starts from a corrected desk wealth SDE (eq. 3) that includes hedge P&L, carry, frictions, collateral benefit, book-funding cost, and default loss. Under CARA utility (eq. 4), the entropic transformation converts the log-expectation into a quadratic running cost, yielding the HJB (eq. 5). The decomposition $v=w_0+u$ separates the clean price (solving a risk-neutral PDE, eq. 6) from the XVA adjustment. The BSDE representation (eq. 9-10) has a Qexp driver with quadratic growth in $Z$, exponential growth in $u$ (from the default term), and linear growth from the book-funding $-f_{t}u$ and real-world drift ${\beta }_t{\partial }_{S}u$. Well-posedness invokes Quenez and Sulem (2013). The perturbative expansion and quadratic reduction yield an explicit Riccati system solved via Bernoulli reduction.

Key Findings

• The discount rate for the P-world XVA stack is $r_{\text{tot}}=f+{\lambda }^P{\zeta }_C$ (eq. 11), not $r_0+{\lambda }^Q{\zeta }_C$ — the Q-world formulation over-provisions by the credit-risk premium factor ${\lambda }^Q/{\lambda }^P$
• The $-\alpha M$ coupling in the Riccati (where $\alpha =f-{\lambda }^P{\zeta }_C$) is essential: omitting it gives incorrect hedge speeds and targets
• The tanh solution sign ($\kappa \tau -\varphi$, not $+\varphi$) ensures convergence to the stable fixed point $M_{\infty }>0$
• The credit-hedge target $q^{\text{tar}}$ decomposes into carry, linear CVA, and funded-book components (eq. 31)
• For ${\lambda }^Q/{\lambda }^P\approx 2$: the P-world credit hedge is half the Q-world level, correctly classified as a capital reserve rather than an operational hedge

Critical Notes

Draft status

The paper has at least 6 unresolved BibTeX references (”?”), making several key claims unverifiable — particularly the well-posedness proof (Theorem 1) which invokes ”?, Thm 2.4” and the source of the desk wealth SDE (eq. 3).

Overstatement of novelty

The claim that the book-funding cost $-f_t{v}_t$ is “absent from prior XVA formulations” requires careful parsing. Burgard & Kjaer (2011, 2013) and Piterbarg (2010) already include funding cost terms in their PDE/replication frameworks. The specific cash-flow mechanism in the wealth SDE may be a legitimate refinement, but the final pricing equations in the existing literature already capture this effect.

Framework-dependent HVA ordering

The claim “HVA is $O(\epsilon )$, not $O(1)$” depends on the specific scaling $\gamma =\epsilon \rho$. Under alternative parametrisations (e.g., fixed risk-aversion with small friction costs), HVA could enter at leading order. This is a property of the expansion, not of HVA itself.

Severe reduction hypotheses

The closed-form results (Sections 6-7) require six strong assumptions: no Lévy jumps (R1), no collateral control $C\equiv 0$ (R2), frozen deterministic coefficients (R3), quadratic frictions (R4), affine delta approximation (R5), and quadratic default surrogate $|u|\ll w_0$ (R6). The paper does not quantify approximation errors from any of these. R6 in particular breaks for large exposures.

No numerical validation

Purely analytical — no Monte Carlo benchmarks, no comparison with existing deep BSDE solvers, no sensitivity analysis. The claim (Remark 3) that the Riccati solution is the policy a neural network would learn is untested.

Empirical uncertainty in ${\lambda }^Q/{\lambda }^P$

The paper’s central quantitative claim (P-world hedge is half of Q-world) rests on ${\lambda }^Q/{\lambda }^P\approx 2\text{-}3$ for investment-grade names. This ratio depends on how the credit spread is decomposed into default probability vs. risk/liquidity premium, which is model-dependent and empirically contested.

Well-posedness citation appears incorrect

Theorem 1 invokes ”?, Thm 2.4” and Quenez & Sulem (2013) “Thm 2.1” for Qexp BSDE well-posedness. However, Quenez & Sulem (2013) contains NO Theorem 2.1 or 2.4 — their Theorem 2.3 covers only Lipschitz drivers, which does NOT include the Qexp driver (quadratic in $Z$, exponential in $u$). The correct reference for Qexp BSDEs with jumps is Gu, Lin & Xu (2023), whose Theorem 3.12 and Remark 3.14 cover exactly this structure. The well-posedness proof needs this citation corrected.

Two different "HVA" concepts in the literature

The paper’s HVA (friction cost, $O(\epsilon )$) and the Crépey school’s HVA (model risk adjustment, $O(1)$) are different quantities sharing the same acronym. Bénézet, Crépey & Essaket (2023) show model-risk HVA = 181 on notional 100. Both trace to Burnett (2021), who introduced HVA for hedging frictions — Crépey reinterpreted it for model risk. The claim “HVA is $O(\epsilon )$, not $O(1)$” is valid for friction-HVA under this paper’s scaling, but risks confusion with the established model-risk HVA literature.

Important References

1. Burgard and Kjaer 2011 — PDE representation for bilateral counterparty risk with funding costs (foundational replication framework)
2. Pham 2009 — Continuous-time stochastic control and optimisation with financial applications (risk-sensitive control, h-transforms)
3. Quenez and Sulem 2013 — BSDEs with jumps, optimisation and applications to dynamic risk measures (well-posedness theory for Qexp BSDEs)

Atomic Notes

• book-funding cost in XVA
• HJB equation for XVA under P
• XVA perturbative expansion
• Riccati system for XVA hedging
• XVA overlay on front-office hedge

---

paper