Conditional Value at Risk (CVaR)

CVaR satisfies all four axioms of a coherent risk measure as defined by Artzner et al.:

• Subadditivity: The risk of a combined portfolio never exceeds the sum of individual risks, encouraging diversification
• Positive Homogeneity: Doubling position sizes doubles the risk measure
• Monotonicity: A portfolio with systematically better outcomes has lower risk
• Translation Invariance: Adding cash reduces risk by exactly that amount

This coherence makes CVaR mathematically superior to VaR, which violates subadditivity and can penalize diversification.

CVaR directly measures the expected loss in the worst α% of outcomes, where α is typically 1% or 5%. For a 95% confidence level:

• VaR answers: "What is the minimum loss in the worst 5% of cases?"
• CVaR answers: "What is the average loss across that entire worst 5% tail?"

This captures the severity of extreme events rather than just their threshold. For a portfolio with a 95% CVaR of $10M, the expected loss during a tail event is $10M, not merely the boundary value.

CVaR can be expressed as a convex function of portfolio weights, enabling efficient optimization using linear programming or convex solvers. Rockafellar and Uryasev (2000) showed that minimizing CVaR is equivalent to solving:

• A linear optimization problem when using scenario-based representations
• A formulation that simultaneously computes VaR and CVaR without requiring non-convex VaR constraints

This convexity guarantees that local minima are global minima, making CVaR-based portfolio construction computationally tractable for large-scale institutional portfolios with thousands of assets.

CVaR is jointly elicitable with VaR, meaning both can be estimated and backtested together using consistent scoring functions. The Fissler-Ziegel loss function provides a strictly consistent scoring rule for the pair (VaR, CVaR).

Practical implications:

• Unlike standalone CVaR, the pair can be evaluated using statistical tests on out-of-sample data
• Enables rigorous model validation by comparing forecasted tail losses against realized exceedances
• Supports regulatory compliance under Basel III, which mandates Expected Shortfall for market risk capital calculations

CVaR is a special case of a spectral risk measure, where risk is computed as a weighted average of quantiles with a specific weighting function. In the spectral framework:

• CVaR assigns equal weight to all losses beyond the VaR threshold and zero weight below
• More general spectral measures can assign smoothly varying weights to different loss levels
• This connects CVaR to the broader theory of distortion risk measures used in insurance and actuarial science

The spectral representation allows CVaR to be expressed as an integral over quantiles, linking it to the Lorenz curve and economic theories of inequality measurement.

Minimizing CVaR is equivalent to solving a distributionally robust optimization problem where the adversary can reshape the loss distribution within a Wasserstein ambiguity set. This duality provides:

• A principled way to handle model uncertainty in loss distributions
• Guarantees that CVaR-optimal strategies perform well even under worst-case distributional shifts
• A bridge between risk management and robust machine learning, where CVaR serves as a loss function for training models resilient to adversarial perturbations

This connection is exploited in adversarial market simulation to train trading agents that are robust to regime changes.

The foundational quantile-based risk measure that CVaR extends. VaR estimates the minimum loss expected at a given confidence level (e.g., 95% or 99%) over a specific time horizon.

• Key Limitation: VaR is not coherent—it fails the sub-additivity property, meaning diversification can appear to increase risk.
• CVaR Relationship: CVaR addresses VaR's blindness to the shape of the tail beyond the threshold by averaging all losses exceeding VaR.
• Calculation: For a 99% VaR of $10M, there is a 1% chance losses exceed $10M, but VaR reveals nothing about whether the excess loss is $10.1M or $50M.

A term used interchangeably with CVaR in regulatory frameworks, particularly Basel III banking standards. Expected Shortfall is defined as the expected loss given that the loss exceeds the VaR threshold.

• Continuous Distributions: ES and CVaR are mathematically identical.
• Discrete Distributions: Subtle definitional differences exist; ES is the strict conditional expectation, while CVaR is the weighted average of VaR and ES.
• Regulatory Adoption: The Basel Committee replaced VaR with ES at the 97.5% confidence level for market risk capital calculations, citing ES's ability to capture tail risk more comprehensively.

A generalization of CVaR that applies a weighting function (risk spectrum) to quantiles of the loss distribution, reflecting the risk manager's subjective aversion to different loss levels.

• CVaR as Special Case: CVaR is a spectral risk measure with a step function spectrum that assigns equal weight to all losses beyond the VaR threshold and zero weight below.
• Exponential Spectrum: Assigns exponentially increasing weight to more extreme losses, encoding stronger tail aversion than CVaR.
• Advantage: Provides a flexible framework for encoding an institution's entire risk appetite curve rather than a single threshold.