Value at Risk carries an implicit promise—that splitting your exposure across multiple positions reduces your aggregate risk. This is submodularity: the whole should be less risky than the sum of its parts. The mathematics works cleanly in calm markets. But empirical analysis of US equity returns reveals something disturbing: VaR systematically violates submodularity during market stress periods, exactly when diversification matters most.
The violation isn't random noise. It's structural. VaR measures risk by asking “what's the worst loss at the 95th percentile?” but this threshold-based approach ignores what happens beyond the cutoff. During volatility spikes, correlations surge and tail events synchronize across assets. The measure that promised to reward diversification suddenly reports that combined positions are riskier than isolated ones—not because the portfolio changed, but because the measure itself is brittle.
Expected Shortfall, by contrast, maintains submodularity even under stress. The difference is simple: it averages losses beyond the threshold rather than stopping at it. This integrative structure preserves the diversification property through coherent distortion of the probability distribution. The math didn't change; the measure did.
This matters beyond finance. Any system that uses threshold-based metrics to encourage distributed structures faces the same trap. Regulatory capital requirements, infrastructure resilience scoring, even performance management—when you measure risk by asking “how bad at the cutoff?” instead of “how bad on average past the cutoff?”, you build a metric that penalizes exactly the behavior it claims to reward, precisely when that behavior is tested. The measure becomes least reliable when most needed.