Hedging under Frictions as Convex Risk Minimization over Neural Policies

Buehler, Gonon, Teichmann, and Wood recast the problem of derivative replication under market frictions as the minimization of a convex, law-invariant risk measure ρ over policies parameterized by feedforward neural networks, providing a computationally tractable framework that accommodates proportional and fixed transaction costs, liquidity constraints, position limits, and market impact simultaneously, without requiring a closed-form pricing formula or an analytically tractable Markov structure for the underlying dynamics. The core theoretical result is a universal approximation theorem for the policy class: for any ε > 0 and any admissible Markovian hedging policy in the classical sense, there exists a neural network of finite width and depth such that its output, evaluated on the observable information set at each rebalancing date, approximates the optimal action to within ε in Lp norm, guaranteeing that the deep hedging approximation class is dense in the set of Markovian strategies as network size grows. The Fenchel–Legendre duality of convex risk measures provides the theoretical scaffolding: for any convex law-invariant ρ satisfying the Fatou property, the dual representation ρ(X) = sup_{Q ∈ M_ρ} {E^Q[−X] − ρ*(Q)} holds, where M_ρ is the risk envelope of ρ and ρ* its conjugate functional. This dual form identifies the optimal hedge δ* as the strategy that simultaneously minimizes the primal objective and achieves the dual supremum — a saddle-point condition that connects deep hedging to the classical theory of equivalent martingale measures and provides an interpretation of the trained policy in terms of the least-favorable measure Q* ∈ M_ρ induced by the risk measure. The indifference price of a contingent claim Φ, defined as the shift p* for which the risk of the hedged position plus the claim equals the risk of the unhedged zero-PnL position, is a further byproduct of the duality structure and can be read off from the optimal policy's terminal distribution. The paper's governance prescription — that the objective function, the friction model, and the training distribution must each be explicitly chosen and documented before the policy is trained — is as important as its technical content. A neural policy is only as disciplined as its training specifications, and in the absence of closed-form guidance, the choices of ρ, C_k, and the simulation measure collectively determine whether the learned strategy is robust or fragile, generalizing or overfitting, auditable or opaque.

The cumulative hedging gain net of frictions is G_T(δ) = Σ_{k=0}^{T-1} [δ_k · ΔS_{k+1} − C_k(δ_k, δ_{k-1})] where δ_k ∈ R^d is the position vector at step k, ΔS_{k+1} = S_{k+1} − S_k the price increment, and C_k(δ_k, δ_{k-1}) a convex cost function penalizing both position changes (transaction costs) and absolute position size (margin or liquidity costs). The policy is a measurable function δ_k = f_θ(I_k) of the information set I_k = σ(S_0, ..., S_k, δ_0, ..., δ_{k-1}), parameterized by neural network weights θ. The deep hedging objective is min_θ ρ(−G_T(f_θ)) where ρ is a convex risk measure; for the entropic risk measure ρ_γ(X) = (1/γ) log E[e^{−γX}] with risk aversion γ > 0, the dual representation gives ρ_γ(X) = sup_{Q: H(Q|P) < ∞} {E^Q[−X] − H(Q|P)/γ} where H(Q|P) is relative entropy, identifying the least-favorable measure Q* as the one of minimum entropy that maximizes expected loss. The indifference price satisfies ρ(−G_T(δ^Φ) + p* − Φ) = ρ(0), where δ^Φ = argmin_δ ρ(−G_T(δ) − Φ) is the optimal hedge of the claim.

The shift from model-based closed-form hedges to risk-measure-minimizing neural policies represents a genuine change in the epistemological status of the hedge: classical delta hedging is defined relative to a specific model and its correctness depends entirely on the model's accuracy, while deep hedging is defined relative to an explicit objective function and friction model whose appropriateness is a design choice that can be documented, audited, and challenged independently of any model assumption about the underlying dynamics. This explicitness about the objective is a governance advantage, even though the resulting policy is a black-box function rather than an analytic formula. The connection to Kramkov's optional decomposition theorem — which characterizes superreplication in the frictionless case — illuminates what the deep hedging framework adds and what it changes. In the frictionless limit C_k → 0 with ρ = CVaR_1 (superreplication), the deep hedging policy converges to the classical superreplicating strategy and the indifference price converges to the superreplication price; but for any finite cost level or any ρ strictly interior to CVaR_1, the policy is an expected-shortfall minimizer rather than a superreplicator, and the resulting position is cheaper but accepts some residual risk. This trade-off between cost and risk is made explicit by the duality structure and can be controlled by varying ρ and C_k, providing a principled way to navigate between the two extremes of expensive full hedges and cheap approximate hedges. The training stability of deep hedging depends critically on the convexity of ρ in the terminal PnL: because G_T(f_θ) is linear in the policy output f_θ at each step, and ρ is convex, the objective ρ(−G_T(f_θ)) is convex in the policy output but generally non-convex in the network parameters θ, so that stochastic gradient descent may converge to a local minimum. For entropic ρ, the objective is log-sum-exp of linear functions of δ, a smooth convex functional in the policy space, and this smoothness accelerates convergence relative to CVaR whose subdifferential is non-trivial.

The desk application of deep hedging is as a benchmark for evaluating the completeness of any friction model used in the standard analytic hedging workflow: by training the neural policy on simulated paths from the desk's current pricing model with the desk's estimated transaction cost function, the resulting policy's PnL distribution provides the best achievable outcome under those assumptions, against which the actual analytic hedge's PnL distribution can be compared. Significant divergence between the neural policy's distribution and the analytic hedge's distribution indicates either that the analytic hedge is suboptimal given the friction model, or that the friction model embedded in the analytic hedge does not correspond to the desk's actual cost structure. The indifference price p*(Φ) computed from the deep hedging framework provides a risk-adjusted bid-ask spread: the ask price p*(Φ) and the bid price p*(−Φ) bracket the frictionless model price from above and below by amounts that grow with both the friction level and the risk aversion parameter γ, providing a quantitative model for spread widening under stressed conditions that can be calibrated from observable market bid-ask data on vanilla options rather than requiring a separate friction-model estimation.

The circularity of training distribution specification is the most severe limitation of deep hedging in practice: the simulation model used to generate the T-period trajectories for training is itself calibrated under assumptions about the underlying dynamics — typically a Markovian diffusion or a jump-diffusion with stationary parameters — and the trained policy f_{θ*} is optimal only with respect to the simulated measure rather than the true data-generating measure. Out-of-distribution regimes — stress events, liquidity crises, correlation breakdowns — that were not well-represented in the training distribution produce policy failures that are not detectable from in-sample calibration diagnostics, and the standard uncertainty quantification techniques based on bootstrap resampling of training paths provide confidence intervals conditional on the simulation model rather than genuine epistemic uncertainty intervals. Non-convexity of the training objective in θ means that convergence guarantees from the convex risk measure theory do not transfer directly to the parameter space, and the degree of suboptimality of a locally-converged θ relative to the global optimum is an unknown quantity that depends on the architecture, the dimension of I_k, and the horizon T in ways not characterized by any existing non-asymptotic theory. For long-horizon hedges with large information sets and deep networks, the optimization landscape is high-dimensional and poorly understood, and practitioners must rely on empirical validation of policy quality rather than theoretical optimality guarantees.