Essay, 2026

Why your stop loss is wrong.

A reader walks through five live simulations of random walks, geometric Brownian motion, and bet sizing. By the end, the question of where to place a stop loss has a precise, and slightly surprising, answer.

Reading time about twenty minutes. Every figure is interactive. Keyboard, screen reader, and reduced motion supported throughout.

Begin reading

Part one

A coin, and a problem.

Suppose you have an edge. A small one, but real. You win 53 percent of your bets. Each bet returns plus or minus one unit. The remaining 47 percent you lose. How much of your capital should you risk on the next wager.

The naive answer is, as much as you can afford to lose. The textbook answer is the Kelly criterion, which we will arrive at in part four. The interesting answer, the one this essay defends, is that almost nobody applies either correctly. Most practitioners use a tool called a stop loss, set at a level that feels safe, and treat that level as if it were a guard rail. It is not a guard rail. It is a permeable membrane that the price keeps pushing through, and the tighter you set it, the more often it will be touched.

To see why, we start as simply as possible. A single coin. A running total. A line.

Figure 1. A biased coin, simulated.

Total 0 · Flips 0

Line chart showing cumulative profit from a sequence of biased coin flips. Starts at zero. Drifts upward over time on average. Spends extended stretches below its previous peak, illustrating drawdown.

Press Flip once to add a coin. Auto flip adds twenty per second. Notice how the line drifts up over time when the probability favours you, but spends long stretches below its previous peak.

This is the central observation of the essay. A process with a positive expectation still spends most of its life below its previous high. That gap between the running maximum and the present value is called the drawdown. Drawdowns are the price of the edge. They are not a malfunction, they are the way the edge arrives.

Part two

What a random walk really looks like.

The coin gave us a discrete, jagged line. Real prices are smoother, but they are not less random. The model practitioners reach for first is geometric Brownian motion, the process behind the Black, Scholes, and Merton option pricing formulas. It assumes a price moves under two forces, a steady drift and a random shock, and that the shocks compound multiplicatively rather than additively.

Formally, the price S evolves as dS = μ S dt + σ S dW, where μ is the annualised drift, σ the annualised volatility, and dW a standard Wiener increment. The continuous form discretises to St+1 = St · exp((μ − σ²/2) Δt + σ √Δt Z), with Z drawn from a standard normal. This is the engine running underneath the next figure.

What you should attend to is not any single path but the shape of the cone the paths trace out. Most of the mass lives close to the median, but a handful of paths drift far above or below. The cone widens proportional to the square root of time, which is the property that makes patience so structurally valuable and impatience so structurally expensive.

Figure 2. Geometric Brownian motion, twenty paths.
Twenty simulated price paths with a 95 percent confidence band drawn from 1,000 sample paths.
Summary statistics for the simulated paths.
StatisticValue
Mean final value…
Median final value…
5th percentile…
95th percentile…

Each line is one simulated future. The shaded band contains the central 95 percent of all such futures, estimated from 1,000 sample paths.

Two facts deserve emphasis. First, raising volatility does not simply widen the band, it lowers the median. The Itô correction term −σ²/2 is why. A pair of strategies with the same drift but different volatilities is not a pair of strategies with the same expected return. Second, lengthening the horizon widens the cone faster than it raises the centre. Patience helps, but only if you can pay for the temporary widening.

Part three

The stop loss illusion.

Now we install a stop loss. A stop is a rule that closes the position at a loss of k percent from entry. The intuition behind it is that of an insurance policy. You give up a small premium of expected return in exchange for the protection of bounded loss. The intuition is half right and half wrong, and the half that is wrong matters more.

What follows is the central simulation of the essay. The same drift, the same volatility, the same horizon, run a thousand times with and without a stop. The figure reports two things: the distribution of final returns, and the probability that the stop is touched at any point along the path.

Figure 3. The cost of a stop loss, 1,000 paths.

With no stop

Mean return
…
Median return
…
P(loss exceeds stop)
…
P(final at or below zero)
…

With stop applied

Mean return
…
Median return
…
P(stop touched)
…
P(final at or below zero)
…

Read the panels side by side. The stopped distribution has a sharp spike at the stop level. The right tail, where the edge actually lives, is truncated. The mean falls. The probability of being touched falls slowly with looser stops and rises quickly with tighter ones.

The reason is not subtle. A path with a positive drift and high volatility has a substantial chance of trading down through any nearby threshold before the drift has had time to assert itself. The square root of time scaling of dispersion guarantees this. Tightening the stop pulls the threshold closer, which raises the probability of touching it, which truncates the upper tail of the realised distribution. The premium you thought you were paying is rarely the premium you actually pay.

This does not mean stops are always wrong. It means a stop is a wager on path, not on outcome, and the wager has to be priced. Three rules of thumb fall out of the simulation above. First, never set a stop tighter than roughly one daily standard deviation, because at that point you are paying premium for noise. Second, if your strategy depends on a positive drift, the stop should be wide enough that the drift has time to dominate the dispersion. Third, if you cannot afford the drawdown that would otherwise have occurred, the right intervention is to size smaller, not to stop sooner.

Part four

Kelly's bet sizing.

If reducing the stop is the wrong lever, what is the right lever. The answer was published by John Kelly in 1956 in the Bell System Technical Journal, and it is the lever everyone reaches for once they understand that drawdowns scale with position size and not with stop tightness. For a bet that pays b to one with probability p, the fraction of capital that maximises the long run growth rate of wealth is

f* = (b p − q) / b, where q = 1 − p.

The formula is short. Its implications are not. Sizing at f* maximises the geometric growth rate. Sizing above f* still has positive expected return, but lower geometric growth, and unbounded drawdowns in any finite horizon. Sizing below f* sacrifices growth in exchange for a quieter equity curve. In practice, most successful practitioners run at one half of Kelly or less, because the assumptions Kelly is built on, in particular that p and b are known, are rarely true and the cost of overestimating them is asymmetric.

Figure 4. Three sizing regimes over 200 sequential bets.

Kelly f* 10%

Three equity curves: one half Kelly, full Kelly, and double Kelly, over 200 sequential bets.

The three lines are one realisation each. Half Kelly trades growth for calm. Full Kelly is the geometric optimum. Double Kelly looks heroic until it does not. Resample a few times and watch how rarely double Kelly is the right choice.

A second observation that does not always make it into the textbooks. Kelly assumes you can place exactly the right fraction. The moment leverage is involved, or the moment the universe of bets is correlated, the formula in the form above is wrong. A portfolio Kelly exists, but it requires the covariance matrix of returns and the vector of expected excesses, both of which are estimated rather than known. The estimation error matters more than the formula.

Part five

Now build your strategy.

The figure below is the synthesis. You configure a drift, a volatility, a stop, and a sizing fraction. The simulator runs a thousand paths and reports five summary statistics. The aim is not to find the configuration with the highest mean return. The aim is to find a configuration whose drawdown distribution you can live with, given your real horizon and your real psychology. There is no objectively correct answer, only a frontier.

Figure 5. Configure, simulate, judge.

Equity cone

Equity cone showing the central 95 percent of simulated paths.

Summary

Mean return
…
Median return
…
Median max drawdown
…
P(stop touched)
…
P(final at or below zero)
…

One last thing. Run the configuration that matches your real strategy, then look at the maximum drawdown panel. That number is the cost of doing business at your sizing. If it is larger than what you can tolerate without changing your behaviour, the lever to pull is the sizing, not the stop. The stop reduces the realised return without reducing the conditional drawdown by very much. The sizing reduces both, proportionally and predictably.

This is the whole argument. A stop loss feels like a guard rail because it is named like one. It behaves like a tax on the right tail of an otherwise positive process. There are good reasons to pay that tax, mainly when the position is poorly understood or the drawdown is genuinely uninsurable elsewhere. Most of the time, the better instrument is the position size, which you set once at the beginning and revisit rarely, rather than the stop, which you fiddle with daily and which the market notices long before you do.

Colophon

All simulations run client side in vanilla JavaScript. Random numbers are generated with a seeded Mulberry32 PRNG and a Box Muller transform, so any figure you see on screen can be reproduced exactly given its seed. The figures are pure SVG, drawn from the data each time you move a control. Type is set in Fraunces and Inter, with JetBrains Mono for the formulas. The colour palette is checked against WCAG 2.2 AA contrast pairs at every level.

Sources: Hull, Options, Futures and Other Derivatives, 11th edition, on the GBM discretisation. Kelly (1956), A New Interpretation of Information Rate, on the bet sizing formula. Thorp (2006), The Kelly Capital Growth Investment Criterion, on the half Kelly heuristic.

Built as a portfolio piece. The full source is on GitHub. Suggestions are welcome.