VaR gives you one number and a quiet, dangerous omission. “There’s a 99% chance you won’t lose more than $1M tomorrow” sounds reassuring — until you ask the only question that actually matters on a bad day: and on the 1% of days I do cross that line, how bad does it get? VaR shrugs. It draws a line in the sand and refuses to look past it. Expected Shortfall is the answer to exactly that question, and it’s why regulators eventually fired VaR from its top job.
You already know what VaR is and the three ways to compute it (historical, parametric, Monte Carlo). This lesson is about the risk measure that lives past the VaR line — its definition, why it behaves better mathematically, and the worked arithmetic to make it concrete.
Before you read — take a guess
Two portfolios both report a 99% 1-day VaR of exactly $1M. Which statement is guaranteed to be true?
VaR’s blind spot: the cliff edge
Analogy. VaR is a flood gauge that promises the river will stay below the top of the levee on 99 days out of 100. Useful — but on the 100th day it tells you nothing about whether the water laps an inch over the wall or swallows the whole town. The levee height is the threshold; the depth of the flood beyond it is a different measurement entirely, and it’s the one that decides whether you’re inconvenienced or drowned.
Formally, VaR at confidence is a quantile: the loss level you’ll exceed only of the time. It is, by construction, the best of the worst cases — the smallest loss in the tail, not a typical one. Everything deeper in the tail gets summarised as “…and worse,” and that ellipsis is where portfolios go to die.
Here’s the trap in one picture: take two books with identical 99% VaR.
| 99% VaR | Shape of the tail beyond | Average loss on a breach day | |
|---|---|---|---|
| Portfolio A (thin tail) | $1.0M | losses cluster just past $1M | ~$1.1M |
| Portfolio B (fat tail) | $1.0M | rare but enormous losses lurk | ~$3.5M |
Same VaR, wildly different danger. A risk manager staring only at VaR would call A and B equally risky and size them the same. The day the tail bites, the desk running B is the one explaining itself to the board. VaR’s whole blind spot is the conditional severity — and a measure that fixes it has to average over the tail, not just locate its edge.
Expected Shortfall, defined
Analogy. If VaR is “how high is the levee,” Expected Shortfall is “on flood days, how deep is the water on average.” It doesn’t ask whether you breach — it assumes you have, and reports the typical damage once you’re underwater.
Definition. Expected Shortfall at confidence — also called CVaR (Conditional VaR), Conditional Tail Expectation, or Expected Tail Loss — is the average loss across the worst fraction of outcomes. In words: condition on the fact that you’ve breached VaR, then take the mean of those losses.
So 95% ES averages the worst 5% of loss outcomes; 99% ES averages the worst 1%. Where VaR points at the edge of the tail, ES integrates over the whole tail and hands you its centre of mass. That single change — from “edge of the cliff” to “average depth of the canyon below it” — is everything.
The solid line is VaR — the loss quantile at your confidence level. The shaded region to its left is the loss tail, and Expected Shortfall is its average depth. Drag the confidence slider up and the line marches deeper into the tail while ES tracks the mean of everything past it.
Read the chart deliberately: the line is VaR, a single point. The shaded area is the set of breach outcomes, and ES is its average — which is why the ES readout always sits to the left of (worse than) the VaR readout. They are not competing numbers; ES is literally a summary of the region VaR refuses to enter.
Pin down the definition.
Pick the right option for each blank, then check.
Expected Shortfall at confidence c is the loss over the worst fraction of outcomes. It is also known as , and because it averages a region that starts at the VaR line, ES is always VaR.
Worked example: VaR vs ES on a sorted sample
Numbers make it click. Take a $10M book and the 100 worst-to-best daily returns from a historical window. We only need the worst tail. Here are the 5 worst days, as percentage losses:
| Rank (worst first) | Loss (% of book) | Loss in dollars |
|---|---|---|
| 1 | 6.0% | $600k |
| 2 | 5.2% | $520k |
| 3 | 4.8% | $480k |
| 4 | 4.1% | $410k |
| 5 | 3.9% | $390k |
95% VaR — the threshold. With 100 observations, the worst 5% is the worst 5 days, so the 95% VaR is the loss you exceed only 5% of the time: the 5th-worst loss, the least bad member of the tail.
95% ES — the average of the tail. ES averages all five of those worst days, not just the boundary one:
So the same data yields a 95% VaR of $390k but a 95% ES of $480k — about 23% larger. VaR saw only the $390k boundary; ES felt the full weight of the $600k and $520k days dragging the average up. And notice the iron rule this illustrates: because ES averages numbers that are all at least as large as VaR,
The gap between them is a tail-fatness gauge: a thin tail keeps ES barely above VaR; a fat tail blows it wide open.
A $20M book's four worst days (the worst 4% of a 100-day window) lost 8%, 6%, 5%, and 3%. What is the 96% Expected Shortfall in dollars?
Why VaR isn’t “coherent”: subadditivity fails
ES doesn’t just see more of the tail — it’s also mathematically better behaved. To say that precisely we need the idea of a coherent risk measure, defined by four axioms a sensible risk number ought to obey. Keep them brief:
- Monotonicity. If portfolio A always loses at least as much as B, then A is at least as risky. (Worse outcomes → bigger risk number.)
- Sub-additivity. . Merging two books can’t manufacture risk out of thin air — diversification never increases total risk.
- Positive homogeneity. Double the position, double the risk: .
- Translation invariance. Adding a fixed amount of risk-free cash reduces the risk number by exactly that same amount.
The load-bearing one is sub-additivity — and VaR violates it. Here’s the famous counterexample.
Hold two separate digital (all-or-nothing) bonds, A and B. Each pays a small coupon but has a 4% chance of a large default loss, and the two defaults are independent. Compute the 95% VaR of each alone:
- A standalone has only a 4% chance of its big loss — and , so the default outcome sits outside the 95% tail. Its 95% VaR captures only the small no-default outcome: tiny. Same for B: is small.
Now combine them. With two independent 4% defaults, the chance that at least one defaults is — comfortably more than 5%. So a default outcome now does fall inside the combined 95% tail, and the portfolio’s 95% VaR jumps to capture a full default loss:
VaR just declared the diversified book riskier than the two standalone bets summed. That’s not merely odd — it’s backwards. A risk measure that punishes diversification can be gamed: split one ugly position into legal entities until each one’s tail probability ducks under the threshold, and VaR cheerfully reports less risk. ES is provably sub-additive (it’s a coherent measure), so it can never produce this paradox — diversify, and ES goes down or stays flat, exactly as intuition demands.
The diversification paradox
VaR can report that splitting one risky book into two lowers total risk, or that combining two books raises it past the sum of their parts. Both are violations of sub-additivity — and both reward the wrong behaviour, letting a desk “diversify” its VaR down while the real tail risk sits untouched. Expected Shortfall is sub-additive by construction, so merging books never inflates ES. When a risk number tells you diversification made things worse, the number is broken, not the diversification.
Match each coherence axiom to what it guarantees.
Pick a term, then click its definition.
Which statement about Expected Shortfall vs VaR is a TRAP — i.e. it sounds right but is actually false?
Why regulators switched
This isn’t just theory — it rewrote banking regulation. Under the Basel FRTB (Fundamental Review of the Trading Book), the market-risk capital charge moved away from 99% VaR and over to 97.5% Expected Shortfall. The committee said the quiet part out loud: VaR “does not capture the tail risk” and isn’t coherent, while ES sees the tail’s shape and rewards diversification properly.
Why 97.5% ES specifically, rather than matching the old 99% number? Calibration. For a normal distribution, 97.5% ES ≈ 99% VaR — so the switch was roughly capital-neutral for well-behaved, thin-tailed books, avoiding a disruptive jump in required capital on day one.
The teeth show up off-normal. For a fat-tailed book, 97.5% ES is strictly tougher than 99% VaR, because ES integrates the heavy tail that VaR steps over. So the regime is gentle on plain-vanilla portfolios and bites exactly where it should: the books stuffed with tail risk that the old VaR charge waved through.
Why 97.5% ES ≈ 99% VaR for a normal
For a standard normal, the 99% VaR sits at the quantile. The 97.5% ES averages everything beyond the quantile, and that conditional mean works out to about standard deviations into the tail. The two land almost on top of each other — which is exactly why Basel could swap the measure without re-scaling every bank’s capital overnight. The divergence only appears once the distribution stops being normal: add fat tails, and the 97.5% ES pushes deeper than the 99% VaR, capturing risk the old charge missed.
The catch: ES is harder to backtest
ES isn’t a free win. Its one genuine weakness is backtesting. Checking a VaR model is clean: you just count how often realised losses breached the VaR line and compare against the expected breach rate — a simple, robust, count-based test (the next lesson’s subject).
ES is a conditional mean of rare events, and that’s a slipperier thing to validate. You’re not counting crossings; you’re estimating the average magnitude of losses on the handful of days you breached — and a handful is all you get, almost by definition. With so few tail observations, the estimate is noisy and a clean pass/fail test is genuinely hard to construct. This backtesting difficulty is the main practical objection to ES, and it’s the reason VaR never fully left the building: regulators set capital with ES but still lean on VaR’s tidy breach count to police whether the underlying model is any good. The two measures end up partners, not rivals.
Big picture
Expected Shortfall — the whole picture
- Expected Shortfall (CVaR)
- What it is
- Average loss in the worst (1−c) of outcomes
- E[L | L ≥ VaR]: mean of the tail past VaR
- Aliases: CVaR, Conditional VaR, Expected Tail Loss
- ES ≥ VaR always
- VaR's blind spot
- VaR is the edge of the tail, not its depth
- Same VaR can hide wildly different tails
- ES sees tail shape; VaR does not
- Coherence
- Four axioms: monotonic, sub-additive, homogeneous, translation-invariant
- VaR fails sub-additivity
- VaR can punish diversification
- ES is coherent: diversify and ES falls or holds
- Regulation
- Basel FRTB: 99% VaR to 97.5% ES
- 97.5% ES ≈ 99% VaR for a normal
- ES strictly tougher on fat tails
- The catch
- Conditional mean of rare events
- Harder to backtest than VaR
- VaR breach count still polices the model
- What it is
Recap: Expected Shortfall
Expected Shortfall at confidence c is best described as:
Check your answer to continue.
Next up — Backtesting VaR (and the limits of the whole framework) — we put these models on trial. We’ll count breaches against expectation with the Kupiec and Christoffersen tests, ask whether the violations cluster (a model that fails all at once is worse than one that fails at random), and confront the honest limits of every VaR-style number: it’s only as good as the history and assumptions you fed it, and the worst day is always the one your sample never saw.