Risk Is Not One Thing

Three things the word risk is doing at once

A household saying "risk" usually means "the chance of losing money", which is a perfectly good place to start and a useless place to finish, because a plan needs a number it can compare against another number. So the word has to be split. The first part is **volatility**: dispersion you can measure. Feed a return series into a standard deviation and you get a number in the same units as the returns, and that number can be compared across strategies, scaled to a horizon, and used to size a position. It is the workhorse of the subject for exactly one reason — it is measurable. The second part is **uncertainty**: what you cannot price at all. Knight drew this line in 1921 and it has never stopped being useful. Volatility describes a coin whose properties you know; uncertainty describes a market where the rules themselves can change — a regulator rewriting settlement, a company restating five years of accounts, a counterparty that turns out to be one person with a concentrated book. Volatility can be estimated from the past. Uncertainty has no sample. Both are risk in ordinary speech and they require different responses: the first is sized against, the second is diversified away from, insured, or simply refused. The third part is the **tail**, and it is what makes the first two dangerous on their own. A normal distribution says a five-sigma day happens once in several thousand years, which is a fine model until the market delivers it twice in a decade. Real return distributions are peaked in the middle and fat in the tails, meaning quiet days are quieter than the model says and bad days are far worse, and they are lopsided as well — equities fall further than they rise, in bursts. So a risk measure built on the middle of the distribution will understate exactly the outcomes that end plans. Same average, different account — Volatile sequence, arithmetic mean: 8.33% a year · Volatile sequence, compound return: 5.53% a year · The gap: 2.80 points — the volatility drag ← · Steady 8.33% over the six years: 2.3 points a year more, compounded The drag is not a fee, a tax or a mistake. It is what compounding does when the returns vary, and it means the honest statement of a strategy return is the geometric figure rather than the average.

Why the average is not what you earn

The arithmetic of it is small. A balance is multiplied by 1 + r each year, so the six annual factors multiply together and the six-year outcome is their product raised to one sixth. Multiplication does not care that the average of the factors looks fine: 1.40 followed by 0.80 is 1.12, not the 1.10 that averaging suggests, and every additional swing adds another of those small shortfalls. The bigger the swings, the wider the gap between the arithmetic mean and the compound return, and the gap is not a rounding error — in the sequence above it is 2.8 points a year. This is the first place the subject becomes uncomfortable, because it means two strategies with identical average returns are not equivalent, and the one with more variance is worth less. It also means that a return figure quoted as an average is a claim about an arithmetic operation and not about anybody account. Where a manager reports 12% a year, the question that follows is whether that is the average of the yearly returns or the compound growth of the money, because on a volatile series the difference can be most of the edge. The other consequence is that the lived experience of a return stream is not its average or its volatility but its **path**: the drawdown in the middle, and how long the account spent underwater. A strategy can average 8% a year and still have spent three years below a previous peak, and the household has to sit through those years with a plan rather than with a forecast. That lived measure — the maximum peak-to-trough fall, and the time to climb back — is the next lesson, and it is where risk stops being a statistic and starts being a decision. What the order does to the same six returns — +40, −20, +30, −25, +15, +10: 1.3814 growth factor · The same returns, reversed: 1.3814 — the product is order-independent ← · What is not order-independent: the drawdown on the way through, and the ending balance of a plan that withdraws A five-sigma day is supposed to happen once in several thousand years. It has happened repeatedly inside a single decade, which is the practical argument against treating a normal distribution as a description of market risk rather than as a convenient approximation.

Volatility clusters, and the calm estimate is the one that fails

There is one empirical fact about volatility that changes how every risk number in this subject should be read: **it is autocorrelated**. Days of high volatility are followed by more of them, and quiet periods persist in the same way. This is not a small statistical curiosity. It means that a risk estimate computed from a calm period is not a neutral estimate of the future — it is a systematically low one, because the calm period is information about a regime that is likely to continue until it does not, and the measurement window contains none of the observations that would raise the estimate. The worst days arrive while the risk model is reporting its smallest numbers. The practical responses are three. Use a window short enough to notice a regime change, and accept that a short window is noisier. Use a method that weights recent observations more heavily, which is what the exponential weighting in the volatility tools of the trading subject does. And separate the level of risk from its tail: a measure of dispersion from a calm window says nothing about the size of the largest move the instrument can make, and those are different questions that are routinely answered with the same number. Clustering also explains why certainty about risk is the wrong goal. A position sized on an estimate of volatility is always sized on a number that is itself uncertain, and the honest way to handle that is to size against a **range** of volatility rather than a point — using the upper end of the plausible range for the position you cannot easily exit, and accepting a smaller position than the calm estimate would permit. The distinction this lesson draws between measurable volatility and unmeasurable uncertainty is not philosophical in a portfolio; it is the reason the position is smaller than the arithmetic suggested. • Volatility is autocorrelated: quiet periods persist and then end abruptly. • Estimates drawn from calm windows are biased low exactly when they are used. • Prefer recent-weighted measures, and treat the level and the tail as separate questions. • Size against a range of volatility, using the high end for positions that are hard to exit.

When volatility is the wrong measure

Volatility is the default risk measure because it captures the whole distribution in one number and behaves well in the arithmetic that follows — it adds across uncorrelated positions, it underlies the portfolio maths, and it is easy to compute. It also treats a gain and a loss of the same size as the same event, which for a distribution that is not symmetric misdescribes what a holder is actually exposed to. The two departures that matter are **skew** and fat tails. Most equity indices have mildly negative skew and heavier tails than a normal distribution allows; the strategies built on selling insurance — option premium collection, carry trades, and every high-win-rate method that collects a small amount most of the time — have strongly negative skew, with a long quiet stretch and an occasional loss that dwarfs the accumulated gains. Their measured volatility looks modest precisely because the rare event is rare *in the sample*, and the sample is exactly the thing that hides the tail. This is not an abstract point: it is the arithmetic resemblance between many products marketed as steady income and the short-volatility complex that failed in a single session in 2018. The alternatives each capture something the standard deviation misses, and each costs something. Semivariance measures only the losses and therefore matches a holder’s experience better, at the price of losing the convenient portfolio maths. Maximum drawdown captures the path and is the number that actually determines whether a strategy survives, but it is sample-dependent and says nothing about what has not happened yet. Expected shortfall describes the average loss beyond a threshold and is a better tail measure than a quantile, though it is estimated from the same sample. And a scenario loss — “what does this book lose if the index falls twenty percent” — is the most useful of all for sizing, with the obvious weakness that it depends on which scenario somebody chose. The practical resolution for an ordinary account is to measure at least two things: a volatility-like number for the ordinary distribution, and a tail or scenario number for the shape. Where the two disagree sharply — low volatility, extreme scenario loss — the disagreement is the finding, and it describes a strategy whose risk sits outside the range that has been observed. Correlation belongs in the same category: it is the least stable input in any variance-based risk system, which means such a system is weakest on exactly the days it is needed. So use volatility for comparability and for the arithmetic that needs it, and size positions from the drawdown you could tolerate and the scenario you could survive. Make the unfamiliar decision — how much you lose if the market does something it has not done in the sample — first, and let the familiar number play its supporting role. • Volatility treats gains and losses alike and hides the left tail of a skewed book. • Short-volatility strategies look calm in the sample because the crash is not in it. • Semivariance, drawdown, expected shortfall and scenario loss each fix part of it. • Where volatility is low and the scenario loss is extreme, believe the scenario. A useful diagnostic on any strategy with a high win rate: plot the distribution of returns and look at the left tail. If the worst loss is many multiples of the average win, the strategy is short volatility whatever it is called, and its size should be set from that loss rather than from its standard deviation.

Sources

• Risk, uncertainty and profit — the measurable/unmeasurable splitFrank Knight, "Risk, Uncertainty and Profit" (1921)
• Volatility drag and the difference between arithmetic and geometric returnsStandard geometric-mean arithmetic; CFA Institute curriculum
• The limits of volatility as a risk measure — fat tails and skewTaleb, "The Black Swan"; Mandelbrot on stable distributions
• Recovery maths and why a deep drawdown is a different riskStandard drawdown arithmetic; Elder, "Come Into My Trading Room"