Chapter 23: Robust Pricing, Hedging, and Decision-Making Under Model Uncertainty

This chapter develops the theory of pricing, hedging, and decision-making when no single probability model can be trusted. Starting from the distinction between risk (known probabilities) and Knightian uncertainty (unknown or ambiguous probabilities), it constructs sets of probability measures as the mathematical representation of model ambiguity, derives model-free pricing bounds via superhedging duality, the Skorokhod embedding problem, and martingale optimal transport, develops robust hedging strategies that perform well across families of models, studies ambiguity-averse preferences and their implications for portfolio choice, builds the nonlinear expectation framework of \(g\)-expectations and second-order BSDEs for pricing under volatility uncertainty, addresses the practical challenges of robust calibration, and concludes with case studies in model failure and the limits of quantitative modeling.

Key Concepts

Knightian Uncertainty distinguishes between risk, where the probability distribution P is known and expectations E[X] = ∫_Ω X(ω) dP(ω) are well-defined, and genuine uncertainty, where the distribution itself is ambiguous. The Ellsberg paradox demonstrates that rational agents systematically violate Savage's subjective expected utility axioms, preferring known risks over ambiguous gambles. The multiple priors representation V(f) = min_P ∈ P E_P[u(f)] (Gilboa-Schmeidler, 1989) provides the axiomatic foundation for decision-making under ambiguity.

Sets of Probability Measures formalize model uncertainty by replacing a single probability measure Q with an ambiguity set Q ⊆ M_1(Ω). The space of probability measures is endowed with the weak topology, where P_n →(w) P if ∫ f dP_n → ∫ f dP for all bounded continuous f. Distance metrics include total variation ||P - Q||_TV = sup_A ∈ F|P(A) - Q(A)| and the Wasserstein distance W_p(μ,ν) = (inf_π ∈ Π(μ,ν) ∫ |x-y|^p dπ)¹/p. Wasserstein balls Q = {Q: W_p(Q, Q_0) ≤ ε} provide distributional robustness with metric structure and finite-sample guarantees.

Misspecification vs. Estimation Error identifies two fundamentally different sources of model uncertainty. Misspecification (P^ ∉ M) means the true data-generating process lies outside the model class; estimation error means the true model belongs to the class but parameters are estimated with finite-sample noise. Under misspecification, the pseudo-true parameter θ_0 = argmin_θ D_KL(P^ || P_θ) provides the best KL-divergence approximation. The distinction determines whether robustness should be structural or statistical.

Model-Free Pricing Bounds derive the tightest price bounds on exotic derivatives using only no-arbitrage and observed vanilla option prices, without specifying a stochastic process. Static arbitrage bounds such as C(K,T) ≥ (S_0 - Ke^-rT)^+ follow directly from no-arbitrage. The fundamental theorem of asset pricing ensures the existence of a risk-neutral measure Q such that discounted prices S_t = E_Q[S_T | F_t] are martingales, and the model-free framework optimizes over all consistent measures.

Superhedging Duality establishes that the model-free upper bound on an exotic payoff Φ equals the cost of the cheapest superhedging portfolio: sup_Q ∈ Q E^Q[Φ] = inf{V_0 : ∃ θ ∈ A with V_T^θ ≥ Φ}. The portfolio value process V_t^θ = θ_t⁰ S_t⁰ + Σ_i=1^d θ_tⁱ S_tⁱ must be self-financing and admissible, connecting dynamic hedging to static optimization over equivalent martingale measures.

Skorokhod Embedding Problem reduces path-dependent option pricing to finding a stopping time τ of Brownian motion such that B_τ has a specified distribution matching the risk-neutral marginal implied by European options. Different embeddings (Root, Rost, Azema-Yor) correspond to different extremal models and yield sharp bounds for lookback, barrier, and variance options.

Hobson's Robust Bounds provide specific sharp model-free bounds for variance options, lookback options, and barrier options derived via the Azema-Yor embedding, proving that certain model-free relationships hold across all consistent models.

Martingale Optimal Transport (MOT) synthesizes optimal transport theory with the martingale constraint. The primal problem sup{E^Q[Φ(S_T_1,...,S_T_n)] : Q ∈ M(μ_1,...,μ_n)} optimizes over all martingale measures consistent with given marginals calibrated to European option surfaces. It extends the classical Kantorovich formulation W_c(μ,ν) = inf_π ∈ Π(μ,ν) ∫ c(x,y) dπ(x,y) by adding the martingale constraint, and the dual problem identifies the optimal semi-static hedge.

Pathwise Hedging constructs portfolios whose P&L bounds hold path-by-path without probabilistic assumptions, working in the path space C([0,T], R_+) with the supremum norm. Follmer's pathwise Ito formula f(S_T) - f(S_0) = ∫_0^T f'(S_t) dS_t + 1/2∫_0^T f''(S_t) d[S]_t enables hedging using only the continuity and finite quadratic variation of price paths, requiring no stochastic model.

Robust Delta-Gamma Hedging minimizes the worst-case hedging error inf_Δ,Γsup_Q ∈ QE^Q[error²] by incorporating second-order (gamma) risk management. The Greeks Δ = ∂ V/∂ S, Γ = ∂² V/∂ S², and V = ∂ V/∂ σ are computed as model-averaged quantities that are more stable than single-model Greeks across the ambiguity set.

Semi-Static Hedging combines static positions in vanilla options V_t = θ_t S_t + Σ_i n_i^C C_i + Σ_j n_j^P P_j + B_t with dynamic trading in the underlying at discrete times, exploiting the Carr-Madan decomposition to replicate exotic payoffs. This approach reduces transaction costs, provides robustness to model misspecification, and decomposes the exotic payoff into a replicable component and a residual bounded by the model-free spread.

Hedging Under Transaction Costs analyzes super-replication prices under proportional transaction costs k, where V^k = V⁰ + O(k²/3) with the Leland correction providing the leading-order adjustment. This framework connects robust hedging to practical market frictions.

Max-Min Expected Utility (Gilboa-Schmeidler) replaces standard expected utility with max_πmin_Q ∈ QE^Q[U(W)], optimizing against the worst-case model. The Ellsberg paradox (f_1 ≻ f_3 and f_2 ≻ f_4 simultaneously) motivates this departure from Savage's framework, and the resulting conservative portfolios explain the equity premium puzzle and home bias.

Multiplier Preferences (Hansen-Sargent) use the criterion max_πmin_Q{E^Q[U(W)] + θ R(Q||Q_0)}, penalizing deviation from the reference model Q_0 via relative entropy rather than imposing hard constraints on Q. The penalty parameter θ controls the degree of robustness.

Entropy Penalization provides the soft-constraint foundation for robust preferences through the Kullback-Leibler divergence D_KL(P||Q) = E_P[log(dP/dQ)]. The variational formula log E_Q[e^θ X] = sup_P ≪ Q{θ E_P[X] - D_KL(P||Q)} connects exponential tilting to worst-case expectations and underpins the duality between entropic penalties and risk-sensitive control.

Risk-Sensitive Control uses the exponential criterion J^RS(x,π;γ) = 1/γlogE[exp(γΣ_t c(X_t,u_t))], where γ > 0 induces risk aversion and γ → 0 recovers the risk-neutral case. The Taylor expansion J^RS ≈ E[C] + γ/2Var(C) shows that risk-sensitive control penalizes both the mean and variance of cumulative cost, providing an equivalent formulation to multiplier preferences through the duality between exponential utilities and entropic penalties.

Robust Portfolio Optimization addresses the sensitivity of mean-variance portfolios w^* = 1/λΣ⁻¹(μ - ν1) to estimation errors by solving max_w min_(μ,Σ) ∈ U{w^Tμ - λ/2w^TΣ w} over uncertainty sets U of plausible parameters. Ellipsoidal, box, and factor-based uncertainty sets lead to tractable conic optimization reformulations.

Black-Litterman and Bayesian Robustness provides implicit robustness through Bayesian shrinkage: combining market equilibrium (prior) with investor views (likelihood) via μ_BL = [(τΣ)⁻¹ + P'Ω⁻¹P]⁻¹[(τΣ)⁻¹π + P'Ω⁻¹q], which shrinks extreme mean estimates toward equilibrium and yields more stable portfolios.

Model Uncertainty in Asset Allocation encompasses factor model selection (CAPM vs. multi-factor), distributional assumptions (Gaussian R ~ N(μ,Σ) vs. Student-t R ~ t_ν(μ,Σ) vs. mixture of normals), return predictability, and regime dynamics. Bayesian model averaging and distributionally robust optimization provide principled approaches to portfolio construction when the correct structural specification is ambiguous.

Dynamic Consistency requires that robust preferences satisfy a recursive structure so that today's optimal plan remains optimal when the future arrives: a_t^(ω) = argmax_a_t V_t(a_t, σ^_t+1:T | F_t = ω). Under standard expected utility with Bayesian updating, dynamic consistency is automatic via the tower property, but ambiguity aversion and non-expected utility models may fail this requirement, leading to time-inconsistent behavior.

Uncertain Volatility Models (Avellaneda-Levy-Paras, 1995) consider volatility constrained to a band σ ≤ σ_t ≤ σ without specifying its dynamics. The super-replication price satisfies the fully nonlinear Black-Scholes-Barenblatt PDE ∂_t V + 1/2σ²(Γ)S²Γ + rSV_S - rV = 0, where the worst-case volatility σ²(Γ) = σ²1_Γ ≤ 0 + σ²1_Γ > 0 is selected at each point based on the sign of gamma.

g-Expectations (Peng) define a nonlinear expectation E_g[X] = Y_0 where (Y_t, Z_t) solves the BSDE dY_t = -g(t, Y_t, Z_t) dt + Z_t dW_t with terminal condition Y_T = X. The generator g encodes the pricing rule; g = 0 recovers classical linear expectation. Properties include monotonicity, constant preservation, and a nonlinear tower property that ensures dynamic consistency of the associated risk measure.

Sublinear Expectations (Peng) define E[X] = sup_Q ∈ QE^Q[X], a coherent nonlinear expectation under which the canonical process is "G-Brownian motion" with uncertain volatility. This framework provides the mathematical foundation for worst-case pricing across an ambiguity set.

Second-Order BSDEs (2BSDEs) (Soner-Touzi-Zhang) generalize classical BSDEs by allowing uncertain quadratic variation a ≤ σ_tσ_t^T ≤ a. The 2BSDE Y_t = ξ + ∫_t^T F(s,Y_s,Z_s,Γ_s) ds - ∫_t^T Z_s dW_s - ∫_t^T tr[Γ_s d⟨ W⟩_s] introduces a second-order process Γ_t whose interaction with the uncertain quadratic variation produces fully nonlinear equations, providing probabilistic representations for second-order nonlinear PDEs and pathwise super-replication prices.

Worst-Case Calibration seeks parameters that ensure acceptable performance under adverse conditions by defining an acceptable parameter set Θ_ε = {θ: CalibrationError(θ) ≤ ε} and computing worst-case prices V^worst = min_θ ∈ Θ_ε V(θ). The min-max hedging problem min_Δmax_θ ∈ Θ_εHedgingError(Δ,θ) finds hedges that minimize worst-case error over all acceptable models.

Stability vs. Fit Trade-Off addresses the tension that models fitting current market data extremely well often have unstable parameters. Parameter sensitivity δθ ≈ -H⁻¹J^Tδ C reveals that ill-conditioned Hessians amplify small data perturbations into large parameter changes. Tikhonov regularization θ_λ = argmin_θ{Σ_i[C_i^market - C_i^model(θ)]² + λ||θ - θ_0||²} and other regularization approaches balance calibration accuracy against parameter stability.

Confidence Sets for Models provide rigorous statistical quantification of parameter uncertainty. A (1-α) confidence region C_n satisfies P_θ(θ ∈ C_n) ≥ 1-α. Likelihood-based regions use Wilks' theorem Λ_n(θ_0) →(d) χ²_p to construct C_n = {θ: 2[ℓ(θ) - ℓ(θ)] ≤ χ²_p,1-α}, while Wald regions form ellipsoids C_n^W = {θ: n(θ-θ)^TI(θ-θ) ≤ χ²_p,1-α} centered at the MLE.

When Models Fail examines historical episodes where model assumptions diverged from reality. LTCM (1998) demonstrated that correlation assumptions ρ_ijⁿormal ≈ 0.2 collapsed to ρ_ij^crisis ≈ 0.8 during the Russian default, with leverage amplifying losses by a factor L√1+(n-1)Δρ. The taxonomy of failure modes includes tail risk underestimation, correlation instability, liquidity evaporation, and feedback effects.

Case Studies in Model Failure analyze LTCM (1998, convergence trade collapse under 25:1 leverage), Barings Bank (1995, operational risk and rogue trading), the London Whale (2012, VaR model misuse on $100B credit derivatives), and the 2008 financial crisis (Gaussian copula failure in CDO pricing, housing correlation underestimation, and systemic liquidity collapse). Recurring themes include correlation breakdown in stress, leverage amplification of model error, and the gap between model-assumed and actual liquidity.

Human Judgment vs. Automation examines the complementary strengths of quantitative models (consistency, speed, scalability, emotional neutrality) and human judgment (pattern recognition in novel situations, contextual understanding, adaptability to regime changes). Cognitive biases including overconfidence, anchoring, and recency bias affect human decision-making, while models lack adaptability to genuinely novel situations outside their training data.

Open Problems in Financial Mathematics include the volatility surface puzzle (no model jointly captures smile, skew, term structure, and dynamics), rough volatility modeling with Hurst parameter H ≈ 0.1 implying E[|logσ_t+Δ - logσ_t|²] ~ Δ²H, the joint calibration problem (fitting vanillas, variance swaps, and VIX options simultaneously), extending robust methods to high-dimensional settings, developing computationally tractable MOT for realistic derivative books, and reconciling statistical and risk-neutral uncertainty.

Role in the Book

Robust pricing provides the conceptual counterpoint to the model-specific frameworks developed throughout the book. The superhedging duality connects to the fundamental theorem of asset pricing (Chapter 1), the uncertain volatility model extends the Black-Scholes PDE (Chapter 6), \(g\)-expectations generalize the Feynman-Kac formula (Chapter 5), robust portfolio optimization complements mean-variance theory (Chapter 4), and the model risk analysis deepens the calibration framework (Chapter 17) and risk management chapter (Chapter 22). The martingale optimal transport framework provides the deepest connection between probability theory and financial mathematics, while the case studies in model failure ground the theoretical developments in practical lessons about the limits of quantitative modeling.

---