Expected Shortfall Beyond VaR

VaR answers “how bad is a bad day?” with a single quantile — the loss you won’t exceed 99% of the time. But it has a notorious blind spot you’ve already glimpsed: it tells you where the tail begins and nothing about how deep it goes. A $10M VaR is consistent with a worst-case of $11M or $500M; VaR can’t tell them apart. This lesson fixes that with Expected Shortfall (ES, a.k.a. CVaR) — the average loss given you’re in the tail — and shows why ES isn’t just “a bigger number” but a better-behaved one: it satisfies the coherence axioms, including subadditivity (it respects diversification), which VaR can violate. This is the number Basel moved bank capital onto, and now you’ll see exactly why.

VaR’s blind spot: a line, not a depth

Analogy. VaR is a flood gauge that tells you the water will reach the second step 99 days out of 100 — but says nothing about whether the hundredth day brings ankle-deep water or a wall that swallows the house. It marks the threshold of disaster, then goes silent on its magnitude.

Definition recap. The Value-at-Risk at confidence $c$ is the loss quantile: $\text{VaR}_c =$ the smallest loss $\ell$ such that $P(\text{Loss} \le \ell) = c$. At $c = 99\%$ it’s the loss you exceed only 1% of the time. By construction, it is the boundary of the tail — and it reports a single point on the distribution, full stop.

Why that’s dangerous. Everything beyond VaR — the entire shape and depth of the tail — is invisible to it. Two portfolios can share a VaR while one has a gentle tail and the other a cliff. Worse, VaR’s blindness invites manipulation: a desk can sell far-out-of-the-money options or take on hidden catastrophe exposure that leaves the 99th percentile untouched while loading the 99.9th percentile with ruin. VaR will bless the position. This isn’t hypothetical — it’s the structural reason a VaR-only risk regime can quietly accumulate tail bombs.

Expected Shortfall: averaging the disaster

Analogy. If VaR tells you the flood will reach the second step, ES tells you the average depth of water on the days it does flood. It stops marking the threshold and starts measuring the catastrophe itself.

Definition. The Expected Shortfall at confidence $c$ is the average loss conditional on being in the tail beyond VaR: $\text{ES}_c = E[\,\text{Loss} \mid \text{Loss} \ge \text{VaR}_c\,].$ Equivalently, it’s the average of all the VaRs at higher confidence levels: $\text{ES}_c = \frac{1}{1-c}\int_c^1 \text{VaR}_u\, du.$ That second form is the key intuition: ES doesn’t read one point off the tail — it integrates over the whole tail, weighting in every worse-than-VaR outcome. So ES is always at least as large as VaR (and strictly larger whenever the tail has any depth), because an average of things all $\ge$ VaR can’t be below VaR.

Names. ES is also called Conditional VaR (CVaR), Tail VaR (TVaR), or expected tail loss. They’re the same object for continuous loss distributions.

Worked example — normal case. For a standard normal loss, the closed forms are $\text{VaR}_c = z_c$ and $\text{ES}_c = \phi(z_c)/(1-c)$, where $\phi$ is the normal density and $z_c$ the quantile. At $c = 97.5\%$: $z_{0.975} = 1.96$, $\phi(1.96) \approx 0.0584$, so $\text{ES}_{0.975} = \frac{0.0584}{0.025} \approx 2.34.$ So a 97.5% ES of 2.34σ is roughly the same magnitude as a 99% VaR of 2.33σ. This is not a coincidence — Basel III deliberately swapped the old 99% VaR for a 97.5% ES precisely because, under normality, they line up, so the switch isn’t a stealth capital hike but a shape-aware upgrade: identical for thin tails, automatically more conservative for fat ones.

Coherence: why ES is the “right” kind of risk measure

There’s a deeper reason to prefer ES than “it sees more of the tail.” A risk measure should obey a few common-sense axioms — and ES does while VaR famously fails one.

Definition — a coherent risk measure (Artzner, Delbaen, Eber, Heath) satisfies four axioms:

1. Monotonicity: if portfolio A always loses at least as much as B, then $\rho(A) \ge \rho(B)$. (Worse outcomes → more risk.)
2. Translation invariance: adding $$k$ of cash reduces the risk by exactly $$k$. (Cash is a buffer.)
3. Positive homogeneity: doubling every position doubles the risk: $\rho(\lambda A) = \lambda\rho(A)$. (Scale linearly.)
4. Subadditivity: $\rho(A + B) \le \rho(A) + \rho(B)$. Diversification can only help — the risk of the combined book never exceeds the sum of the parts.

ES satisfies all four. VaR can violate subadditivity — and that’s the deal-breaker.

Why subadditivity matters. Subadditivity is the mathematical statement of “diversification reduces risk.” If a risk measure can increase when you combine two books — punishing diversification — it’s pathological: it could push a firm to split a diversified portfolio to lower its reported risk, the opposite of prudence. A risk measure that doesn’t respect diversification can’t be used to aggregate risk across desks, which is the whole job of a firm-wide measure.

A worked subadditivity violation

Let’s see VaR fail with the cleanest possible example — two independent defaultable bonds.

Setup. Each bond pays a small coupon but has a 4% chance of a large default loss. Specifically, each bond independently loses $100 with probability 4% and gains $2 with probability 96%. Consider 95% VaR (so we care about the worst 5%).

Each bond alone. The probability of the $100 loss is 4%, which is less than the 5% tail. So the 95% quantile of each single bond’s loss falls in the no-default region: the 95% VaR of one bond is just $-2$ dollars (a $2 gain, i.e. negative loss). VaR says each bond alone is essentially riskless at 95%.

The portfolio of both. Now hold both. The chance that at least one defaults is $1 - 0.96^2 = 1 - 0.9216 = 0.0784$, i.e. 7.84% — now above the 5% tail. So the 95% quantile lands in the default region: with one default you lose $100 - 2 = 98$ dollars. The 95% VaR of the combined book is about $98.

The violation. Sum of individual VaRs: $(-2) + (-2) = -4$ dollars (a gain). Combined VaR: $+98$ dollars (a loss). So $\text{VaR}(A+B) = 98 \;>\; \text{VaR}(A) + \text{VaR}(B) = -4.$ Diversifying made VaR explode. That’s the subadditivity violation — VaR punished a perfectly sensible diversification. ES would not do this: because it averages over the whole tail (including both-default scenarios), it correctly registers each bond’s default risk even when it sits just outside the 5% cut, and the combined ES stays below the sum. This single example is why theorists distrust VaR and regulators migrated to ES.

EVT-based Expected Shortfall

ES truly shines when fused with the GPD tail from the last lesson, because then you can compute ES far past your data.

The neat GPD result. If the tail above the threshold is GPD with shape $\xi$ (and $\xi < 1$, so the mean exists), then for a confidence level in the tail, $\frac{\text{ES}_c}{\text{VaR}_c} = \frac{1}{1-\xi} + \frac{\beta - \xi u}{(1-\xi)\,\text{VaR}_c}.$ The dominant term is the clean ratio $\dfrac{1}{1-\xi}$: for a pure power-law tail, ES is just VaR scaled up by $1/(1-\xi)$. The heavier the tail (bigger $\xi$), the bigger the multiplier — exactly the right behavior.

Worked example. Take the EVT fit from last lesson: $\xi = 0.25$, and a 99.9% VaR we computed at about 7.3% (a $7.3M loss on $100M). The leading-order ES multiplier is $\frac{1}{1-\xi} = \frac{1}{1-0.25} = \frac{1}{0.75} \approx 1.33.$ So $\text{ES}_{0.999} \approx 1.33 \times 7.3\% \approx 9.7\%$, i.e. about a $9.7M average loss given you’re in that worst-0.1% region. Notice the gap: VaR said $7.3M, but the typical loss when you breach is $9.7M — a third worse. If $\xi$ were a scarier 0.5, the multiplier would be $1/0.5 = 2$, doubling VaR to $14.6M. The tail index directly controls how much worse ES is than VaR — a one-number summary of “how bad is bad, on average.”

Next up — copulas and tail dependence — we leave single-asset tails and confront how assets crash together. Correlation, it turns out, is a fair-weather friend: assets can look mildly linked day-to-day yet plunge in lockstep during a crisis. Copulas separate “how each asset behaves” from “how they move together,” and tail dependence captures the “everything goes to one” clustering that destroys diversification exactly when you need it.