Minimum CVaR Optimization
Portfolio optimization that minimizes Conditional Value-at-Risk, a coherent risk measure that captures the expected loss in the tail of the return distribution.
Overview
Conditional Value-at-Risk (CVaR), also known as Expected Shortfall (ES), measures the expected loss conditional on the loss exceeding the Value-at-Risk (VaR) threshold. Unlike VaR, which only tells us the minimum loss at a given confidence level, CVaR tells us how bad things get on average when they are bad. This makes CVaR a far more informative risk measure for portfolio management.
CVaR is a coherent risk measure in the sense of Artzner et al. (1999), satisfying the properties of monotonicity, sub-additivity, positive homogeneity, and translation invariance. Sub-additivity is particularly important as it ensures that diversification never increases risk -- a property that VaR famously violates in certain cases.
The seminal contribution of Rockafellar and Uryasev (2000) showed that CVaR minimization can be reformulated as a linear program, making it computationally tractable even for large portfolios with many scenarios.
Risk Measure Definitions
Value-at-Risk (VaR)
For a loss random variable and confidence level (typically 0.95 or 0.99), VaR is the -quantile of the loss distribution:
Conditional Value-at-Risk (CVaR)
CVaR at confidence level is the expected loss given that the loss exceeds the VaR threshold:
For a continuous loss distribution, CVaR can equivalently be written as the average of all VaR values at levels above :
This integral representation shows that CVaR is a weighted average of quantiles in the tail, providing a complete picture of tail risk rather than a single point estimate.
Portfolio Optimization Formulation
General Formulation
The minimum CVaR portfolio solves the following optimization problem:
where is the portfolio loss (negative return) and the optimization is over the weight vector . Additional constraints such as target return, sector limits, or turnover bounds can be added.
Linear Programming Reformulation
The key insight of Rockafellar and Uryasev (2000) is that CVaR can be minimized through the auxiliary function :
where is an auxiliary variable that, at the optimum, equals the VaR. Minimizing jointly over and yields the minimum CVaR portfolio.
Sample-Based LP Formulation
Given historical return scenarios , the problem becomes a linear program by introducing auxiliary variables :
Each captures the excess loss beyond the VaR threshold for scenario . The objective averages these excesses over the worst fraction of scenarios. This LP has variables and constraints (plus bound constraints), and can be solved efficiently with standard LP solvers.
Relationship Between VaR and CVaR
CVaR always exceeds or equals VaR at the same confidence level:
The gap between CVaR and VaR depends on the shape of the tail distribution. For normally distributed returns, the relationship is deterministic:
where and are the standard normal PDF and CDF. For heavy-tailed distributions (which are typical in financial markets), the CVaR-VaR gap is larger, reflecting greater tail risk.
Advantages
- Coherent risk measure: Satisfies sub-additivity, ensuring that diversification is always rewarded.
- Tail sensitivity: Captures the full shape of the loss tail, not just a single quantile.
- LP tractable: Can be formulated as a linear program, enabling efficient computation for large-scale problems.
- Distribution-free: Works directly with historical scenarios or simulated samples, requiring no distributional assumptions.
- Regulatory relevance: Basel III/IV regulations mandate Expected Shortfall for market risk capital calculations.
Limitations
- Sample dependence: Estimation quality depends heavily on the number of scenarios; tail estimation requires many observations.
- Non-smooth: The sample CVaR function is piecewise linear and not differentiable everywhere, requiring LP or specialized solvers.
- Conservative: May produce overly conservative portfolios by focusing exclusively on tail losses.
- Confidence level sensitivity: Results can be sensitive to the choice of , and there is no universal standard for its selection.
References
- Rockafellar, R.T. & Uryasev, S. (2000). "Optimization of Conditional Value-at-Risk." Journal of Risk, 2(3), 21-42.
- Krokhmal, P., Palmquist, J., & Uryasev, S. (2002). "Portfolio Optimization with Conditional Value-at-Risk Objective and Constraints." Journal of Risk, 4(2), 43-68.
- Artzner, P., Delbaen, F., Eber, J.M., & Heath, D. (1999). "Coherent Measures of Risk." Mathematical Finance, 9(3), 203-228.