WhitmanTrading

What Is a Deviation Risk Measure?

Deviation risk measure is a formal class of statistics that score how far outcomes spread, not how much can be lost. The defining property is that adding a constant to every outcome leaves the number unchanged, which is exactly why it cannot answer a question about loss.

There is a formal difference between a statistic that says how scattered outcomes are and one that says how much money is at stake. Almost every risk figure quoted casually is the first kind.

Defined exactly

A price series with the spread of outcomes marked.
A deviation risk measure scores spread, not loss. Illustrative chart - not real market data.

A deviation measure scores dispersion. It takes a distribution of outcomes and returns one number describing how far they travel from their own centre.

A steady series with dispersion around a mean.
Standard deviation is the familiar example. Illustrative chart - not real market data.

The class has four required properties. It is never negative and is strictly positive for anything that varies. It scales proportionally: double every outcome and the measure doubles. It is subadditive, so combining two things never scores worse than the sum of their parts — which is the property that makes diversification show up in the arithmetic.

A rising series where a constant shift leaves the statistic fixed.
Add a constant and the number does not move. Illustrative chart - not real market data.

And the fourth is the one that matters. Adding the same constant to every outcome leaves the measure completely unchanged.

A falling series illustrating the defining invariance.
Which is the defining property of the class. Illustrative chart - not real market data.

Standard deviation satisfies all four, which is why it is the example everybody already knows.

The property that defines the class

A choppy series indifferent to the level of outcomes.
It is blind to the level of the outcome. Illustrative chart - not real market data.

Constant-shift invariance sounds technical and is not. Take a strategy’s monthly returns. Add ten percentage points to every single month. The deviation measure returns exactly the same number it did before.

A slow series where the statistic cannot describe loss.
So it cannot say how much you can lose. Illustrative chart - not real market data.

Now subtract twenty from every month instead. Still the same number. A statistic that behaves identically for a strategy making money and one bleeding steadily is not describing loss, and no amount of context makes it start.

A calm series showing distance travelled from average.
It says how far from average things travel. Illustrative chart - not real market data.

What it is describing is uncertainty about the average, wherever that average happens to sit. That is genuinely useful information and it is a different question from the one most people think they are asking when they quote it.

A worked example

Average true range is a deviation-style statistic, and this site’s series has it measured: the median fourteen-period reading is 0.5994 and the ninetieth percentile is 0.7954.

Divide one by the other and the ratio is about 1.33. A busy stretch runs roughly a third wider than a typical one, which is the dispersion figure doing its job.

A falling series with a stop level marked.
A stop fills where the market is. Illustrative chart - not real market data.

Now add a hundred to every price in the series. Every bar’s range is unchanged, because a range is a difference — so the ATR readings are identical, and so is the ratio. The statistic did not notice.

Drawdown does notice. The deepest decline on the same series is 3.76% and the longest stretch below a prior peak runs 73 bars, and both of those are measured against a level rather than a spread.

That is the whole distinction in two figures from one file. One pair of numbers survives a shift untouched; the other pair is defined by position relative to a high-water mark.

Why it is not the same as a risk measure

A risk measure in the formal sense answers a capital question. How much would have to be set aside for this position to be acceptable — and adding cash to a position reduces that requirement one for one.

Which is the opposite behaviour. A risk measure moves when you add a constant; a deviation measure is defined by not moving. They are separate families with separate axioms, and the two are routinely used as if they were interchangeable.

They are formally linked, though. Applying a deviation measure to a series after its mean has been subtracted produces a corresponding risk measure, and the correspondence runs both ways for a wide class of them.

So the practical rule is simple. A deviation number tells you nothing on its own. Pair it with a mean — which is exactly what the Sharpe ratio does — or it cannot rank anything.

Deviation does not mean symmetric

Standard deviation treats an outsized gain and an outsized loss identically, and that symmetry is the usual complaint against it.

But symmetry is not required by the definition. Lower semideviation, which counts only outcomes below the mean, satisfies every one of the four properties and is a deviation measure in full standing.

So does the deviation form of conditional value at risk, built from the average of the worst tail measured against the mean rather than against zero.

Which means the fix for asymmetry is inside the class, not outside it. The Sortino ratio works precisely by swapping one deviation measure for a one-sided one and changing nothing else.

The original data

On this site’s shared series: median bar range 0.493, ninetieth percentile 1.101, largest single bar 2.338. Direction runs average 2.01 bars with a longest of 11.

Divide the largest bar by the median and it is about 4.7 times. A dispersion statistic fitted to the typical bar treats that day as extraordinarily unlikely, and it happened inside this sample.

A candlestick chart annotated with the cost of a round trip.
A round trip costs a share of a bar. Illustrative chart - not real market data.

And the runs figure says the outcomes are not independent. An average run of 2.01 bars with a longest of 11 is the clustering that makes a spread number fitted on the whole sample understate what a bad stretch looks like while it is happening.

A price series with volume shown beneath.
Volume and price measure different things. Illustrative chart - not real market data.

What it is genuinely for

Comparing the variability of things on the same scale. Two strategies, same market, same period — the deviation figure says which one delivers its average more consistently.

Feeding an optimiser. Subadditivity is what lets a portfolio construction routine find combinations scoring better than their parts, and it is a property of the measure rather than of the market.

Sitting in the denominator of a performance ratio, where the mean supplies the direction and the deviation supplies the scale.

And flagging a regime change. A deviation figure that doubles has said something real about conditions even though it has said nothing about direction.

When it fails

The characteristic failure is treating the number as a loss estimate and sizing from it. A position is built so that a two-deviation move is survivable, on a series where the largest bar measured 4.7 times the median and direction runs reach 11 bars. Those two facts together mean the distribution has a fat tail and the bad outcomes arrive consecutively, neither of which a single dispersion figure encodes. The position survives every ordinary week and is decided by the one stretch the statistic was structurally incapable of describing, because the statistic was never measuring loss in the first place.

A candlestick series with a gap through a level.
A gap skips the level entirely. Illustrative chart - not real market data.

A second failure is ranking strategies on deviation alone, which is blind to whether either of them made money.

A third is assuming symmetry is part of the definition, and discarding the whole class over a property only some members have.

A fourth is comparing figures across sampling frequencies without rescaling, which makes a daily number and a monthly one look like a real difference.

A declining series cut short at a sizing decision.
The spread is unchanged. Is the exposure? Illustrative chart - not real market data.

And a fifth is quoting it as “risk” in front of someone who will hear that word as downside risk and act accordingly.

Standard deviation covers the member of the class everyone already uses. Downside risk covers the question people usually meant to ask. And distortion risk measure covers the other formal family, built to answer it.

What I actually do

The thing that made this click for me was the constant-shift property. A number that does not move when you add ten percent to every single outcome is plainly not measuring how much you can lose — and standard deviation, which everybody quotes as a risk figure, is exactly that number.

— Michael Whitman

This page is educational, not financial advice. Test every idea on your own charts before risking money.