Distortion Risk Measures

An insurer pricing a policy against hurricane damage cannot just use the raw historical probability of a hurricane; if it did, the premium would only cover the average payout, with nothing left over for the insurer's own risk of ruin in a bad year. What insurers have done for a century instead is quietly inflate the probability of bad outcomes before averaging, computing a premium as if disasters were more likely than the data says, then charging accordingly. That trick, formalized, is exactly what a distortion risk measure does to a portfolio's loss distribution.

The analogy before any symbols

A worried parent checking a weather forecast that says "10% chance of a storm" might mentally round that up to "let's plan as if it's 25% likely," not because she disbelieves the forecast, but because she wants extra caution baked into her planning specifically around the bad scenario. She has not changed the weather; she has changed how much weight the bad scenario gets in her own decision-making. A distortion risk measure formalizes exactly this instinct: take the real probabilities, deliberately distort them to inflate the bad tail, then compute an ordinary average under the distorted, more pessimistic probabilities.

The mechanics

Let $F(x) = P(\text{loss} \leq x)$ be the true cumulative distribution of losses, so $1-F(x)$ is the true probability of losing more than $x$. A distortion function $g: [0,1]\to[0,1]$ is a non-decreasing function with $g(0)=0$ and $g(1)=1$, applied to the survival probability:

$\text{distorted survival probability} = g\big(1-F(x)\big).$

If $g$ is concave (curves downward), it inflates every survival probability except at the endpoints, $g(p) \geq p$ for $p \in (0,1)$, meaning "the true 5% chance of losing more than $x$" becomes, say, a distorted 12% chance, deliberately overstating how likely bad outcomes are. The distortion risk measure is then the expected loss computed under this inflated probability:

$\rho_g = \int_0^\infty g\big(1-F(x)\big)\, dx.$

In words: instead of reweighting individual scenarios directly (the spectral-measure approach), warp the probability scale itself so that the region near "loss exceeds a large threshold" gets stretched out and treated as more likely than it truly is, then take an ordinary expectation on that warped scale. VaR is the distortion measure with $g$ a step function (jumps from 0 to 1 exactly at $1-\alpha$); expected shortfall is the distortion measure with $g(p) = \min(p/(1-\alpha), 1)$, a function that ramps up linearly and then caps at 1. Every spectral risk measure, see Spectral Risk Measures, corresponds to exactly one distortion function, they are two equivalent ways of describing the same family of ideas: reweight scenarios directly, or warp the probabilities feeding into an ordinary expectation.

Apply a concave distortion to this curve's tail probabilities and the effective picture used for pricing or risk-budgeting stretches the left tail outward, treating a true 1-in-100 loss as though it were closer to 1-in-30, without changing a single actual data point.

Worked example: pricing insurance with a distortion

A policy pays $1m with true probability 2%, otherwise $0. Fair (undistorted) expected payout is $0.02 \times 1{,}000{,}000 = 20{,}000$, i.e. $20,000, which is what the insurer would charge with zero margin for its own risk. Using the Wang transform, a common distortion, $g(p) = \Phi(\Phi^{-1}(p) + \lambda)$ for a risk-aversion parameter $\lambda$, with $\lambda = 0.5$: $\Phi^{-1}(0.02) = -2.054$, so $g(0.02) = \Phi(-2.054+0.5) = \Phi(-1.554) \approx 0.060$. The distorted probability is 6.0%, three times the true 2%. Distorted expected payout, the premium: $0.060 \times 1{,}000{,}000 = 60{,}000$, i.e. $60,000, three times the actuarially fair price. The extra $40,000 is the loading that compensates the insurer for taking on tail risk, generated entirely by the probability distortion, not by any change to the actual claim distribution.

Worked example: comparing two distortions on the same book

A trading book's loss distribution has a 1% true chance of a $5m loss and a 4% true chance of a $1m loss, otherwise breakeven. Under a mild distortion ($g$ close to linear), the distorted probabilities might become 1.8% and 5.5%, giving a distorted expected loss of $0.018\times5{,}000{,}000 + 0.055\times1{,}000{,}000 = 90{,}000 + 55{,}000 = 145{,}000$, i.e. $145,000. Under a more aggressive distortion tuned to punish extreme tail events harder, the 1% event might distort to 4.5% while the 4% event only distorts to 7%, giving $0.045\times5{,}000{,}000 + 0.07\times1{,}000{,}000 = 225{,}000 + 70{,}000 = 295{,}000$, i.e. $295,000, more than double, because the more aggressive distortion specifically inflates the rarer, larger loss far more than the more common, smaller one. Two "reasonable" distortion choices can disagree by 2x on the same book, purely from how aggressively each inflates the far tail relative to the near tail.

What this means in practice

Distortion risk measures are the standard theoretical language in actuarial pricing (insurance premiums are essentially always computed with some distortion built in, whether acknowledged or not) and they show up in financial risk management wherever VaR or ES is discussed as a special case of a broader family, useful for understanding why those two specific measures behave the way they do rather than as a distinct practical tool most trading desks compute directly.