Spectral Risk Measures

Expected shortfall averages every outcome beyond the VaR threshold with equal weight: the 96th percentile loss counts exactly as much as the worst loss in the sample. That is a specific, somewhat arbitrary choice. A risk manager who is far more worried about catastrophic tail scenarios than merely-bad ones might reasonably want to weight the very worst outcomes more heavily than the merely-bad ones within that same tail. Spectral risk measures generalize expected shortfall to allow exactly that, a chosen weighting scheme across the whole loss distribution, subject to one rule that keeps the measure sensible.

The analogy before any symbols

A teacher grading a group project could give every team member's individual score equal weight in the team's final grade, that is expected shortfall's flat weighting inside the tail. Or she could decide that the weakest performer's score should count for more, since a team's real vulnerability is its weakest link, deliberately weighting low scores more heavily than high ones. She could not, however, sensibly weight a higher individual score more than a lower one when computing a "weakest link" grade, that would defeat the purpose. A spectral risk measure is this teacher's grading scheme applied to a portfolio's possible losses: choose any weighting across outcomes, as long as worse outcomes never get less weight than better ones.

The mechanics

Order possible portfolio losses from best to worst, and let $q(u)$ denote the loss at the $u$-th percentile (so $u$ near 0 is a mild outcome, $u$ near 1 is the worst). A spectral risk measure is a weighted average of losses across every percentile:

$M_\phi = \int_0^1 \phi(u)\, q(u)\, du,$

where $\phi(u)$, the risk spectrum, is a weighting function satisfying three conditions: it integrates to 1 (weights form a valid average), it is non-negative, and it is non-decreasing in $u$ ($\phi(u_2) \geq \phi(u_1)$ whenever $u_2 > u_1$). In words: build the risk number as a weighted average of outcomes across the whole range of possibilities, and the weighting is only required to never favor a better outcome over a worse one, everything else is a free design choice. That monotonicity condition is what makes a spectral risk measure automatically coherent, in particular it always rewards diversification, unlike VaR.

Expected shortfall is the special case where $\phi(u) = \frac{1}{1-\alpha}$ for $u > \alpha$ and $0$ otherwise, a flat weight on the worst $(1-\alpha)$ fraction of outcomes and zero everywhere else, a step function. A genuinely spectral measure instead uses a smoothly increasing $\phi(u)$, so even outcomes inside the "good" region get some small weight, and outcomes in the tail get progressively more weight the worse they are, rather than a single flat plateau.

Imagine sliding a weighting curve like the one above over this distribution's left tail: a spectral measure lets the far-left, most catastrophic sliver of outcomes pull the risk number up more than a slightly-less-catastrophic sliver just next to it, something a flat-weighted average cannot do.

Worked example: an exponential spectrum on five scenarios

Five equally-likely scenarios, losses ordered worst to best: $500k, $300k, $150k, $80k, −$20k (a small gain). Expected shortfall at the 80% level (worst 1 of 5) is just $500k, ignoring the other four entirely. A spectral measure using weights that decay exponentially toward better outcomes, say normalized weights $0.42, 0.26, 0.16, 0.10, 0.06$ on the five scenarios from worst to best (each roughly 60% of the previous, then renormalized to sum to 1), gives $M_\phi = 0.42(500) + 0.26(300) + 0.16(150) + 0.10(80) + 0.06(-20) = 210 + 78 + 24 + 8 - 1.2 = 318.8$, i.e. $318.8k. Unlike VaR's $500k-or-nothing threshold view or ES's flat-tail average, this number reflects every scenario, weighted by how bad it is, capturing more information from the full distribution while still respecting that the worst outcome should never be underweighted relative to a better one.

Worked example: checking the coherence payoff

Two independent desks each have a 5% chance of a $1m loss, otherwise breakeven. VaR at 95% is $0 for each individually, since the 5% tail sits exactly at the confidence boundary, but the combined book has roughly a 9.75% chance of at least one desk losing (assuming independence: $1-(0.95)^2$), so combined 95% VaR jumps to $1m, more than the sum of the individual VaRs, exactly the non-subadditivity that makes plain VaR fail as a coherent measure. Any spectral risk measure with a strictly increasing $\phi$, including expected shortfall, instead gives a combined risk number that is provably no greater than the sum of the two individual numbers, correctly rewarding the (partial) diversification between the two desks rather than penalizing it.

What this means in practice

Spectral risk measures let a risk manager encode a specific attitude toward catastrophic tail risk directly into the risk number, choosing $\phi$ to reflect, say, how a CRRA utility function in Utility Functions and Risk Aversion would weight outcomes, rather than being stuck with either VaR's single quantile or ES's flat tail-average. In practice expected shortfall remains the dominant choice because it is simple, well-understood, and satisfies regulatory backtesting requirements, but spectral measures show up in academic tail-risk research and in bespoke internal risk frameworks at firms wanting a more calibrated view of extreme scenarios.