Your Win Rate Has a Margin of Error

Your journal says 61.8%.

It says it in a nice big number, probably in green, probably at the top of the dashboard. Thirty-four trades, twenty-one winners. And the number feels solid, because it has a decimal place, and things with decimal places feel measured.

That 61.8% isn't your win rate, though. It's your best guess at your win rate, built on a fairly small pile of trades. The real number, the one that would show up if you could somehow trade this setup ten thousand times under the same conditions, sits somewhere in a range around it. With 34 trades, that range is wide enough to drive a truck through.

How wide? Roughly 45% to 79%.

Most of us have stared at a win rate like that and quietly decided the setup works. I have. The math says we were reading a number that couldn't tell us apart from a coin.

What a confidence interval actually is

Skip the textbook version. The trader version goes like this.

Flip a fair coin 34 times and you don't get exactly 17 heads. You get 15, or 20, or 22. The coin hasn't changed. The sample just wobbles. That wobble is called sampling noise, and it's completely normal.

Your trades wobble the same way. If your setup genuinely wins 55% of the time, a sample of 34 trades might show 62%. Or 48%. Or 41%. Same setup, same edge, different draws.

A confidence interval is the honest way of saying: given the wobble, these are the true win rates that could plausibly have produced what I saw. It's a band, not a promise.

When people say "95% confidence interval," they mean a band built so that, over the long run, this kind of calculation captures the true value about 95 times out of 100. It's a statement about the method rather than about your particular setup, and it's close enough to intuition to be useful.

The arithmetic, by hand

You need three things: your number of wins, your number of trades, and a calculator that does square roots.

Call your observed win rate p (as a decimal) and your trade count n.

Step 1. Multiply p by (1 minus p). Step 2. Divide that by n. Step 3. Take the square root. This is your standard error. Step 4. Double it. That doubled number is your margin of error. Doubling is where the phrase "two sigma" comes from, and it's just the convention for drawing a band wide enough to cover roughly 95% of the plausible outcomes. Step 5. Add and subtract it from your win rate.

Let's run the 34-trade example properly.

• p = 0.618, n = 34
• 0.618 × 0.382 = 0.2361
• 0.2361 ÷ 34 = 0.006944
• √0.006944 = 0.0833, so about 8.3 percentage points
• Double it: 16.7 percentage points
• 61.8% ± 16.7% = 45.1% to 78.5%

That interval contains 50%. Which means the data you've collected is genuinely consistent with a setup that loses slightly more than half its trades, and also consistent with one that wins nearly four out of five. Both stories fit.

You haven't measured an edge. You've measured a very loud silence.

The napkin rule worth memorizing

You won't always want to run five steps. This shortcut is close enough for trading and easy to hold in your head.

Near a 50% win rate, your two-sigma margin of error is about 100 divided by the square root of your trade count, in percentage points.

That falls straight out of the formula. At p = 0.5, the p × (1−p) term is 0.25, its square root is 0.5, and doubling it gives you 1 over √n. In percent, that's 100 over √n.

Read that table slowly, because it explains an entire genre of trader frustration.

At 100 trades, a 55% win rate still can't be told apart from a coin flip. At 200 trades the band's lower edge is still 48%, so a coin flip is not yet ruled out. You need something like 400 trades before a 55% win rate becomes a statement rather than a hunch. It's the same reason there is no magic number of trades that validates a setup: the sample you need depends entirely on how big the effect is that you're trying to see, and a five-point edge over a coin is a small effect.

One thing in your favour: the table above is the widest case. The further your win rate sits from 50%, the smaller the p × (1−p) term gets, and the tighter your band. An 80% win rate over 50 trades gives you a margin of 11.3 points instead of 14.1, which helps a little without changing the picture much.

When 50% is not your line anyway

This is where a lot of win-rate talk goes wrong, and where I think most journals quietly mislead people.

50% only matters as a threshold if you win and lose the same amount. Almost nobody does. Your actual break-even win rate depends on your payoff:

Break-even win rate = 1 ÷ (R + 1)

R is just your risk on the trade. One R is the amount you lose when your stop gets hit, so a 2R winner made you twice what you were risking, and a 1R loser cost you exactly what you planned to lose.

If your average winner is 2R and your average loser is 1R, you break even at 1 ÷ 3 = 33.3%. At 3R, you break even at 25%. At 0.5R, you need 66.7% just to stay flat.

So the question is never "does my interval contain 50%." It's does my interval contain my break-even win rate.

Worked example. Say you have 120 trades of a swing setup, 54 winners, so 45%. Your average winner runs 2R, your average loser is a clean 1R.

• 0.45 × 0.55 = 0.2475
• ÷ 120 = 0.0020625
• √ = 0.0454, so 4.5 points
• Doubled: 9.1 points
• Band: 35.9% to 54.1%

Your break-even point is 33.3%. Your lower bound is 35.9%. The interval clears it, but not by much.

Translate that into expectancy and it gets very honest very fast. At your observed 45%, expectancy is (0.45 × 2) − (0.55 × 1) = +0.35R. At the pessimistic end of your band, 35.9%, it's (0.359 × 2) − (0.641 × 1) = +0.08R.

So the fair summary reads: this setup is probably making about a third of an R per trade, and the realistic bad case is roughly break-even after costs. That's a completely different sentence from "45% win rate, 2R average, +0.35R expectancy," and it's the more truthful one. If you want the full walkthrough of how those pieces fit together, start with what expectancy actually measures.

This is the instinct we built into EdgeFlow, just applied to R instead of win rate: discovered edges get ranked by the cautious lower end of their expectancy range rather than by the headline average, so a setup only climbs the list when its pessimistic case still looks worth trading.

Where this arithmetic breaks

I'd rather you know the limits than trust the formula too far.

It gets ugly at the extremes. The simple method above (statisticians call it the Wald interval) misbehaves when your win rate is very high or very low and your sample is small. Nine wins out of ten gives you 90% ± 19 points, which implies an upper bound above 100%. Nonsense. There's a cheap fix that works surprisingly well: pretend you took four extra trades, two winners and two losers, then run the same five steps. For 9 of 10, that becomes 11 of 14 = 78.6%, with a margin of ±21.9 points, so about 57% to 100%. Still wide, but no longer impossible.

It assumes your trades are independent and comparable. If you changed your stop placement halfway through the sample, or the sample spans a volatility regime shift, or you took six correlated positions on the same news event, the interval is measuring something that doesn't exist. Consistent tagging matters more here than the arithmetic does.

It says nothing about the size of your wins. A win rate interval is blind to payoff. You can have a beautifully tight win rate band sitting on top of an expectancy that is held up entirely by two outlier trades. Win rate is one of several things a journal should be showing you, and on its own it's one of the weaker ones. The full set of metrics that actually matter tells a much more complete story.

It gets much too optimistic if you went looking. This is the big one. If you sliced your journal into twelve subgroups, found that London-session continuations won 68%, and then computed an interval around that 68%, the interval is lying to you. You didn't test one hypothesis, you tested twelve and kept the winner. That's a different statistical problem, and it's the main reason a hot subgroup and a lucky streak are so hard to tell apart.

Nobody can hand you a formula that fixes that last one. The only real defence is discovering on one chunk of your data and confirming on a later chunk you haven't looked at yet.

So what do you do with a wide interval

Not nothing. That's the mistake people make when they first meet this idea. They see the width, decide statistics has ruined trading, and go back to reading the point estimate.

The useful move is smaller than that. Start writing your win rate as a range, in your own notes, every time. Not just "61.8%." Write "61.8%, range 45–79%, n=34." It takes four seconds and it changes how you think about the number, because you can't pretend any more that it's precise.

Then behave accordingly. A setup whose interval clears its break-even point is worth sizing up on. A setup whose interval straddles break-even is worth continuing to trade at small size purely to collect more trades, since more trades is the only thing that narrows the band. And a setup whose upper bound sits below break-even is worth stopping, which is genuinely useful information and the case people forget exists.

A win rate whose interval still spans your break-even point isn't a finding. It's a placeholder, your journal saying "ask me again in 60 trades."

That's not a failure of your strategy. It's what a small sample looks like when you're honest about it, and being early in the data is a normal place to be. The traders who blow up on a 34-trade sample usually aren't the ones without an edge. They're the ones who never checked how wide the number was before they leaned on it.

If you want to see your own numbers with the uncertainty attached rather than hidden, that's roughly the whole idea behind how EdgeFlow reports expectancy.