The variance risk premium: why selling volatility usually works, and why that is the problem

August 14, 2026

Implied volatility exceeds subsequent realized volatility on average, across markets and decades. That gap is real and it is harvestable. It also produces the most misleading equity curves in finance, for a structural reason worth understanding before you read another win rate.

There is a persistent, well documented gap in options markets: implied volatility tends to exceed the realized volatility that follows. Sell optionality systematically and you collect that difference.

This is the variance risk premium, and it is probably the most reliably documented premium in derivatives. It is also the source of more blown-up accounts than any other idea in retail options, and both facts have the same cause.

Why the premium exists

The gap is not a market inefficiency waiting to be arbitraged away. It is the price of insurance.

Someone holding a large equity portfolio wants protection against a crash. They are willing to pay more than the statistically fair value for it, because the loss they are insuring against is not just large in expectation, it arrives at the worst possible time, alongside everything else in the portfolio falling together.

The seller collects that premium in exchange for accepting the risk. Both sides are behaving rationally. The premium is compensation, not a mistake, which is exactly why it has not been competed away in the decades since it was first documented.

There is a second reason the gap persists: selling volatility requires capital and the willingness to hold a position through a drawdown that can be sudden and severe. That is a genuine constraint, and constraints are what allow premia to survive.

The payoff shape, which is the whole story

A volatility-selling strategy has a characteristic distribution:

Most periods produce a small gain. The option expires worthless or is

bought back cheaper. Nothing happens, which is the point.

Occasionally a period produces a very large loss. Much larger than the

typical gain, and often larger than many typical gains combined.

That shape is not a flaw in a particular implementation. It is what the premium is compensating you for. You are being paid a steady amount to accept a rare, large, correlated loss. A strategy with this payoff that had no large losses would not be earning a risk premium at all.

The consequence for every number you will read about it

This distribution breaks the metrics people habitually use.

Win rate becomes almost meaningless, and specifically misleading. A volatility seller can be right 85% or 90% of the time and still lose money over a full cycle, because the arithmetic of the payoff has nothing to do with the frequency of the payoff. A high win rate here is a description of the strategy's structure, not evidence of its quality. Any strategy that sells optionality must have a high win rate. It tells you nothing about whether this one is any good.

Average return over a short window is unreliable. If the sample does not contain a stress event, the average describes only the quiet regime.

Sharpe ratio flatters it. Sharpe uses standard deviation, which treats a distribution of many small gains as low-risk. It is low-variance until the tail arrives, and the tail is not in the variance until it happens.

Maximum drawdown in the sample is the wrong number unless you know the sample contained the worst case. It usually did not.

The one question that separates a real test from a bad one

Does the backtest period contain a genuine volatility shock?

A volatility-selling strategy tested from 2012 to 2017, or 2023 to 2025, has not been tested. It has been observed during the conditions in which it works. The result will look outstanding and will tell you nothing about the thing you actually need to know, which is what happens on the bad day.

February 2018 is the standard illustration and it is worth knowing precisely. A widely held short-volatility exchange-traded product lost the overwhelming majority of its value in a single session and was subsequently terminated. It had performed well for years. The years of good performance and the single catastrophic session were the same strategy behaving exactly as designed.

What honest evaluation looks like

Expectancy, not win rate. Average profit per trade including the losses, which requires the losses to be in the sample.

The worst single outcome, stated explicitly. Not the average loss, not the standard deviation. The worst one, and an account of whether the sample could plausibly contain the true worst case.

Position sizing derived from the tail, not the average. The relevant question is not what a typical loss costs, but whether a five-sigma session ends the account.

Explicit treatment of correlation. In a real shock everything sold moves together. A book of twenty short-premium positions across twenty tickers is not twenty independent bets. It is one bet, wearing twenty tickets.

Where this leaves the premium itself

The variance risk premium is real. It has been documented across equity indices, currencies and commodities, over long periods, by independent researchers. It is one of the more robust findings in the field.

Real and harvestable are not the same as safe or easy. Harvesting it means being the insurer, and insurers make steady money for long stretches and occasionally pay out enormously. An insurer who has not modelled the payout has not modelled the business.

The reason this matters beyond options is that the same shape appears whenever a strategy sells optionality in any form, including some that do not look like options at all: carry trades, some mean-reversion systems, and anything whose losing trades are allowed to stay open while winners are closed.

If a track record shows a very high win rate and small average wins, the first

question is not how it did it. The first question is where the large loss is,

and whether the sample was long enough to contain one.