Four Rules Every Risk Number Should Obey: Artzner and the Birth of Coherent Risk

Most papers in finance propose something. This one asked a question that nobody had bothered to ask properly: what should a risk number even mean?

In the 1990s, banks were measuring risk with Value-at-Risk. Regulators were building capital rules on top of it. Nobody had stopped to write down the basic properties any risk measure ought to satisfy. Artzner, Delbaen, Eber and Heath did exactly that, and in doing so they discovered that the industry's favorite measure fails one of the most important tests.

The problem: everyone had a risk number, nobody had a definition of risk

Think about how strange the situation was. A bank would say "our risk is 40 million dollars." That number was produced by a specific formula, VaR, chosen mostly because it was intuitive and JP Morgan had popularized it. But why that formula? What makes a formula a legitimate measure of risk, as opposed to just some number computed from a distribution?

Artzner and coauthors approached it the way a mathematician approaches any concept that people use fluently but have never defined. They started from purpose. A risk measure, they argued, should tell you something specific and actionable: how much extra cash would you have to add to this position to make it acceptable? Risk, in this framing, is measured in dollars of required capital buffer. That is a definition a regulator can use, and it is a definition an internal risk committee can use.

From that anchor, they wrote down four axioms. Any measure satisfying all four, they called coherent.

The key idea via analogy: four things a bathroom scale must do

Suppose you want to check whether a device really measures weight. You would insist on some basic sanity properties. If you put on a heavier object, the number must go up. If you weigh two bags separately and then together, the combined reading should be the sum. And so on. If a device violates those, it is not a scale, whatever it says on the box.

The four coherence axioms are that kind of sanity check, applied to risk.

1. Monotonicity. If portfolio A always does at least as well as portfolio B in every possible future state, then A cannot be riskier than B. This is as basic as it sounds, and it is the one axiom nobody disputes.

2. Translation invariance. If you add a pile of risk-free cash to a portfolio, the required buffer should drop by exactly that amount. Add 10 million of cash, and the risk number falls by 10 million. This is what pins the risk measure to a dollar interpretation.

3. Positive homogeneity. If you double the size of every position, you double the risk. Scale in, scale out.

4. Subadditivity. This is the important one, and the one that caused all the trouble. The risk of a combined portfolio must be no greater than the sum of the risks of its parts. Merge two trading desks and the merged entity should not be riskier than the two desks standing apart, because diversification can only help, or at worst do nothing.

Subadditivity is the mathematical statement of the single most fundamental idea in all of investing: diversification does not hurt.

The bombshell: VaR is not coherent

Value-at-Risk satisfies the first three axioms. It fails subadditivity.

You can construct real, not especially exotic, portfolios where combining two positions produces a VaR that is larger than the sum of the two individual VaRs. It happens naturally with things like out-of-the-money options, defaultable bonds, and any payoff with a lumpy, discontinuous loss profile. The intuition: VaR only looks at a single quantile, a threshold point in the distribution. When you combine two positions, the losses that were each individually just past the 99th percentile can pile up inside the combined tail in a way the threshold measure handles perversely.

The consequences are not academic:

• It penalizes diversification. A risk system built on VaR can literally report that a diversified book is riskier than a concentrated one, which is the opposite of what a risk system exists to do.
• It breaks risk aggregation. Add up the VaRs of your desks and you have no guarantee that this bounds the firm's VaR. The number does not aggregate.
• It creates an incentive to hide risk. A trader can arrange a position with a small VaR and a catastrophic tail: sell far out-of-the-money options, collect premium, and report a low risk number because the disaster sits just beyond the quantile the measure looks at. VaR literally cannot see it.

The paper also showed the constructive side. Coherent measures have a clean characterization: any coherent risk measure can be expressed as the worst-case expected loss across a set of scenarios (what they called generalized scenarios). That connects abstract axioms directly to the very practical world of stress testing, which was reassuring: the way good risk managers already thought, in terms of "what if these scenarios happened," turns out to be the mathematically coherent way.

And they pointed toward the fix. The tail conditional expectation, the average loss given that you are in the bad tail, is the natural coherent alternative. That idea, refined by Acerbi and Tasche and made computable by Rockafellar and Uryasev, is what the world now calls expected shortfall or CVaR.

Why it mattered

• It gave risk management a theoretical spine. Before, risk measures were chosen by convention and convenience. After, there was a standard: does your measure satisfy the axioms, and if not, exactly which sanity property are you giving up?
• It set the intellectual agenda for two decades. Convex risk measures, spectral risk measures, distortion risk measures, the entire modern literature is a response to and extension of this paper.
• It changed regulation. The eventual decision by the Basel Committee to move the market risk framework from VaR toward expected shortfall traces directly back to the incoherence of VaR that this paper exposed. It took years, and a financial crisis, but the argument that started here won.

The honest limitations

• Positive homogeneity is arguable. The axiom says doubling a position doubles its risk. In real markets, doubling a position more than doubles risk, because a bigger position is harder to exit, especially in a crisis. This is liquidity risk, and coherence as originally axiomatized does not capture it. Later work on convex risk measures relaxes this axiom precisely to allow for it.
• Coherence is a minimum bar, not a ranking. Many measures are coherent. The axioms tell you which measures are disqualified, not which of the survivors is best for your problem.
• A coherent measure with bad inputs is still garbage. Expected shortfall computed from a model that has never seen a crash is not safer than VaR computed from the same model. The axioms discipline the formula, not the data.
• The measure is only as good as the distribution. All of this assumes you have a probability distribution over future outcomes. The deepest risks are the ones your distribution does not contain, and no axiom can save you from that.

The one-line takeaway

Artzner, Delbaen, Eber and Heath asked what properties a risk number must satisfy to deserve the name, wrote down four, and showed that Value-at-Risk fails the one that says diversification cannot make you riskier, which is why the profession has been migrating away from it ever since.