Expected Shortfall
Back in
what VaR measures,
we flagged a gap and promised to come back to it: a 99% VaR tells you a threshold that's rarely
crossed, but says nothing about how bad things get on the rare days it is crossed. Two
portfolios can share an identical VaR while one has a merely-bad worst day and the other could
end the firm. Expected Shortfall (also called Conditional VaR or
CVaR) is the fix: instead of asking "where's the cutoff?", it asks
"given that we're past the cutoff, how bad is it on average?"
Definition: the average loss, beyond the cutoff
The c% Expected Shortfall is the average loss conditional on the loss
exceeding the c% VaR:
\text{ES}_c = \mathbb{E}\bigl[\,L \;\big|\; L > \text{VaR}_c\,\bigr].
In a historical-simulation or Monte Carlo sample of M scenarios, this
is entirely mechanical: sort the scenarios worst to best exactly as before, but instead of
reading off one value at rank k=\lceil (1-c)M\rceil, take the
average of all k worst scenarios:
\text{ES}_c = \frac{1}{k}\sum_{t=1}^{k} \text{P\&L}_{(t)}, \qquad k = \lceil (1-c)M \rceil,
where \text{P\&L}_{(1)} is the worst scenario,
\text{P\&L}_{(2)} the second worst, and so on. Because it
averages a whole slice of the tail rather than pinpointing its edge, Expected Shortfall
is always at least as large as VaR at the same confidence level — it can never be smaller.
Worked example: extending the historical simulation
Recall the 20-day historical sample from
historical simulation,
reweighted onto a $10,000,000 portfolio, with worst-to-best returns
-3.4\%, -2.2\%, -1.5\%, -1.1\%, \dots. The 90% VaR needed
k=\lceil 0.10\times 20\rceil = 2, landing on the second-worst day
alone: \text{VaR}_{90} = 0.022 \times \$10{,}000{,}000 = \$220{,}000.
Expected Shortfall instead averages those same two worst days:
\text{ES}_{90} = \frac{(-3.4\%) + (-2.2\%)}{2}\times \$10{,}000{,}000 = -2.8\%\times \$10{,}000{,}000 = \$280{,}000.
\$280{,}000 > \$220{,}000 — exactly as guaranteed. The VaR figure only
ever saw the boundary of the tail; the ES figure looked inside it, and reported a bigger,
more honest number.
Picture it: two lines on the same tail
On the standardized P&L bell curve, VaR is the dashed cutoff line; Expected Shortfall is a
second, solid line sitting further into the tail — the probability-weighted average of
everything past the cutoff. The shaded region is identical to
the one you met before;
now it has two landmarks instead of one.
Why "coherent" matters: VaR can call diversification bad
A risk measure is called coherent if it satisfies four sensible axioms, the most
important of which is subadditivity: combining two positions into one portfolio
should never make the reported risk go up. Diversification should never look like a bad
idea. Remarkably, plain VaR can violate this.
Take two independent bonds, each with a 4% chance of defaulting (a $100 loss) and
a 96% chance of paying in full (no loss). At 95% confidence:
-
Each bond alone: \mathbb{P}(\text{loss} > 0) = 4\% \le 5\%,
so \text{VaR}_{95} = \$0 for either bond by itself.
-
Both bonds together: the chance that at least one defaults is
1 - (0.96)^2 = 7.84\% > 5\%, so the $0 threshold no longer satisfies
the definition — the next possible loss level, $100 (exactly one default), is needed:
\text{VaR}_{95} = \$100 for the combined portfolio.
Sum of the individual VaRs: \$0 + \$0 = \$0. VaR of the combined,
diversified portfolio: \$100. Holding both independent bonds
together looks riskier, by VaR's own arithmetic, than holding either alone — precisely
the failure subadditivity forbids. Expected Shortfall, by contrast, is provably coherent: no
portfolio of independent (or any) positions can ever make ES claim that diversifying increased
risk.
The 2007–09 financial crisis exposed VaR's blind spot in the worst possible way: banks that
reported comfortable VaR numbers right up until the crisis were then hit by losses far beyond
anything their VaR figure had hinted at, because VaR had never been asked to describe the tail
— only to mark its edge. The academic case for something better was already decades old (Philippe
Artzner, Freddy Delbaen, Jean-Marc Eber and David Heath formalized "coherent" risk measures,
with Expected Shortfall as the star example, in a landmark 1999 paper), but it took the crisis
to force regulators' hands. Basel III's Fundamental Review of the Trading Book
(finalized 2016) replaced VaR with Expected Shortfall as the standard measure banks must hold
capital against for market risk — a direct, practical consequence of the subadditivity failure
above.
The single most common misreading of a VaR number is treating it as a worst-case estimate.
VaR tells you a threshold is crossed some fraction of the time — it says absolutely
nothing about how bad things get once it's crossed. A 99% VaR of $1M is exactly as
consistent with a worst day of $1.01M as it is with a worst day of $50M; VaR alone cannot tell
those two firms apart. Expected Shortfall exists specifically to answer the question VaR cannot:
not "is the threshold crossed?" but "how bad is it, on average, when it is?"