CVaR & Expected Shortfall Optimization Lab

1. Coherent Risk Measures & The Subadditivity Dilemma

In landmark research, Artzner, Delbaen, Eber, and Heath (1999) established the formal axiomatic definition of a coherent risk measure \(\rho: \mathcal{X} \to \mathbb{R}\):

• Monotonicity: If \(X \le Y\), then \(\rho(X) \ge \rho(Y)\) (greater portfolio losses require higher capital reserves).
• Translation Invariance: \(\rho(X + c) = \rho(X) - c\) for cash addition \(c\).
• Positive Homogeneity: \(\rho(\lambda X) = \lambda \rho(X)\) for all \(\lambda > 0\).
• Subadditivity: \(\rho(X + Y) \le \rho(X) + \rho(Y)\) (diversification never increases aggregate risk).

Value-at-Risk (VaR) fails the subadditivity condition. Under non-normal or skewed distributions, \(\text{VaR}_\alpha(X + Y) > \text{VaR}_\alpha(X) + \text{VaR}_\alpha(Y)\) can occur. This creates a severe moral hazard in portfolio management: a risk manager optimizing for VaR may concentrate portfolio exposure into short volatility or catastrophic tail-risk assets that keep VaR artificially low while ballooning extreme losses beyond the threshold.

2. The Rockafellar-Uryasev (2000) Auxiliary Formulation

R. Tyrrell Rockafellar and Stanislav Uryasev solved the longstanding computational challenge of minimizing CVaR without relying on non-convex integer programming. For a portfolio with weights \(w \in \mathbb{R}^n\) and loss distribution \(f(w, y) = -w^T r\), they defined the auxiliary function: \[ F_\alpha(w, \zeta) = \zeta + \frac{1}{1 - \alpha} \int_{y \in \mathbb{R}^n} \bigl[ f(w, y) - \zeta \bigr]^+ p(y) \, dy \] where \([t]^+ = \max(t, 0)\) and \(\zeta\) is a real scalar. Rockafellar and Uryasev proved the foundational theorem: \[ \min_{w \in W} \text{CVaR}_\alpha(w) \equiv \min_{w \in W, \zeta \in \mathbb{R}} F_\alpha(w, \zeta) \] Furthermore, the scalar \(\zeta^*\) that minimizes \(F_\alpha(w, \zeta)\) is precisely the \(\text{VaR}_\alpha(w)\) of the portfolio! Under \(S\) Monte Carlo or historical scenarios \(r_1, \dots, r_S\), this continuous expectation approximates as: \[ F_\alpha(w, \zeta) \approx \zeta + \frac{1}{(1 - \alpha) S} \sum_{s=1}^S u_s \quad \text{subject to} \quad u_s \ge -w^T r_s - \zeta, \quad u_s \ge 0 \] This converts the complex non-smooth CVaR minimization problem into a standard, globally solvable Linear Program (LP).

3. Basel III / FRTB 97.5% Expected Shortfall Shift

The Basel Committee on Banking Supervision (BCBS) enacted the Fundamental Review of the Trading Book (FRTB), fundamentally replacing 99% VaR with 97.5% Expected Shortfall (ES). The regulatory rationale: \[ \text{ES}_{97.5\%} = \frac{1}{0.025} \int_{0.975}^1 \text{VaR}_u \, du \] For a standard normal distribution, \(\text{VaR}_{99\%} \approx 2.326 \sigma\) and \(\text{ES}_{97.5\%} \approx 2.338 \sigma\), making the baseline capital charges roughly comparable. However, under empirical financial fat-tailed distributions (such as Student-t or jump diffusion), \(\text{ES}_{97.5\%}\) is substantially higher, compelling banks to reserve capital against tail contagion, liquidity freezes, and asymmetric crashes.

4. Euler Component CVaR Tail Risk Decomposition

Because CVaR is homogeneous of degree one (\(\text{CVaR}(\lambda w) = \lambda \text{CVaR}(w)\)), Euler's theorem allows exact additive attribution: \[ \text{CVaR}_\alpha(w) = \sum_{i=1}^n w_i \frac{\partial \text{CVaR}_\alpha(w)}{\partial w_i} = \sum_{i=1}^n \text{CCVaR}_i \] where the Marginal CVaR is the conditional expectation of asset \(i\)'s losses during portfolio tail disasters: \[ \frac{\partial \text{CVaR}_\alpha(w)}{\partial w_i} = \mathbb{E}\bigl[ -R_i \;\big|\; -w^T R \ge \text{VaR}_\alpha(w) \bigr] \] This reveals hidden systemic risks: an asset may represent only 15% of portfolio capital yet contribute over 50% of the entire portfolio's tail crash severity.