Guide · Part II — The mental tools

From VaR to CVaR

The risk number most funds quote is the height of a fence. The number that matters is how far the ground drops on the other side. Those are different measurements, they have different names, and confusing them is how well-modelled portfolios produce unmodelled disasters.

The problem

VaR — Value at Risk — answers the question “how much could be lost in a normal bad period?” Formally: 95% of the time, losses will not exceed some stated amount. It sounds like a worst case. It is the opposite: it is the edge of normal, the boundary of the ordinary 95%, with total silence about the remaining 5%.

The silence is the flaw. Two portfolios can share an identical VaR while one falls slightly past the fence in a bad month and the other falls two hundred times as far. VaR cannot tell them apart, because it only marks where the tail begins, never how deep it runs. In 2008, every major bank ran a VaR model; the models were roughly right about the 95th percentile and catastrophically silent about the 1st, which arrived far worse than the fence implied.

The insight

CVaR — Conditional Value at Risk, also called Expected Shortfall — asks the question VaR refuses: given that this is one of the worst 5% of outcomes, what is the average loss across those scenarios? It is the mean of the tail, not its edge.

VaR is the fence; CVaR is the average depth of the fall past the fence.

The deeper reason mathematicians insist on CVaR is called coherence. Artzner and colleagues laid out four common-sense axioms a risk measure should obey; the load-bearing one is sub-additivity — combining two portfolios must never show more risk than the two measured separately, because diversification can help but should never mathematically hurt. VaR violates this: merging two unrelated positions can make measured VaR rise, which is nonsense, and means a firm cannot safely add VaR figures across its book. Acerbi and Tasche showed expected shortfall satisfies all four axioms. This is not aesthetic tidiness — it means CVaR aggregates honestly, while stacked-up VaR can lie.

A pair of illustrative portfolios shows what the tail-average catches that the fence misses. Portfolio A posts a calm plus 0.7% every month for thirty months — a beautiful smooth line — then falls 45% in month thirty-one. Portfolio B bounces noisily, up 8% one month, down 7% the next, worst single month minus 15%, but never breaks and always recovers. Most beginners point to the smooth line as the safer one. A’s average loss in its worst outcomes is catastrophic and permanent; B’s is uncomfortable and survivable. The smoothness was the disguise, not the safety.

In plain English

Replace “what is the average return?” with “what is the average loss in the worst months — and is that survivable without selling at the bottom?” That substitution is the entire chapter in one move. For any fund or product, ask for the CVaR or, in plainer clothes, the ten worst months on record, and average them: that is the felt risk. If the ten worst months cannot be produced, something important has been learned anyway — the risk is being hidden, not absent.

The wiggle that can be seen is rarely what ruins anyone; the hole that cannot be seen is. A return series that never wiggles is often a series that cannot wiggle — because the asset is rarely re-priced, or because incoming money and borrowed money are papering over the cracks. Smoothness in a risky world is a claim that should raise the level of scrutiny, not lower it.

Where this breaks

CVaR is estimated from a sample of historical returns, and the whole machine assumes the future tail resembles the past one. Three classes of loss break that assumption outright.

• Custody and counterparty failure — an exchange collapsing, an issuer’s reserves proving fictional, an irrecoverable private key — is a total loss triggered by something other than a price move, appearing in zero return histories because it is not a return but a disappearance.
• Regulatory reclassification or an outright ban is a step-function: an overnight repricing no continuous distribution contains.
• A short, calm sample produces a shallow tail estimate no matter how sophisticated the formula — a risk number is only as deep as the worst event in its data.

The honest hierarchy is therefore three levels, not two. VaR marks the edge of normal. CVaR averages the depth beyond the edge, for risks that are in the sample. And the tail-of-tails — legal, custodial, sovereign — was never in the sample at all, and requires a different defence entirely: the insurance logic of the tail-hedging chapter, and the pre-commitment logic of the discipline layer.

The action is a two-question overlay on every holding. First: can this position lose everything from something that is not a price move? Second: can a regulator or a government reprice it overnight? A yes to either means the position’s true risk is not in any chart it will ever be sold with.