How does downside risk optimization differ from mean-variance?

Question

AcadiFi · Accepted Answer

Downside risk optimization replaces variance with measures that only penalize negative deviations, addressing the critique that investors care more about losses than gains. Key frameworks:

1. Lower Partial Moment (LPM): LPM_n(τ) = E[max(τ - R, 0)^n]. Order n = 2 gives semivariance; n = 1 gives expected shortfall below threshold; n = 0 gives probability of loss. Optimization minimizes LPM at target τ (often risk-free rate or 0).

2. Mean-Semivariance: replace covariance matrix in Markowitz with semicovariance Σ_s = E[(R - μ)_- (R - μ)_-'] where (x)_- = max(-x, 0). Harlow's post-modern portfolio theory formulation. Gives different frontier than mean-variance, typically shifting weights toward low-skewness low-downside-vol assets.

3. CVaR Optimization: minimize Conditional Value at Risk at confidence level α. Rockafellar-Uryasev formulation converts to linear programming using historical scenarios. Widely used for fat-tailed distributions.

4. Value-at-Risk Optimization: minimize quantile loss. Non-convex, difficult to solve; approximated by CVaR in practice.

5. Sortino Ratio Maximization: covered separately—maximize (return - MAR) / downside deviation.

Comparison: mean-variance assumes normality or quadratic utility where variance equals downside variance. Real returns have skew and fat tails. Downside methods better align with investor loss aversion (Kahneman-Tversky) and post-2008 risk management practices. Implementation requires more data for stable downside covariance estimation—typically 5-10 years monthly. Solvers include specialized LP/QP for CVaR and iterative algorithms for semivariance frontiers. Institutional LDI strategies increasingly use downside-focused objectives aligned with funding status.

How does downside risk optimization differ from mean-variance?

Master Level III with our CFA Course

Related Questions

Practice Questions