You have seen the post. Probably more than once.
"1 confluence = 40% win rate. 2 confluences = 55%. 3 confluences = 70%+."
Sometimes it is a carousel. Sometimes it is a table in a course PDF with a green highlight on the bottom row. The numbers shift a little between versions, but the shape is always the same: stack more reasons to take the trade, and your win rate climbs in a tidy staircase.
Quick definition first, because the word gets thrown around loosely. A confluence is just one condition you require before you take a trade. Trend pointing your way. A sweep of yesterday's low. London session. Price at a level you drew earlier. Each of those is one confluence, and "stacking confluences" means demanding several of them at the same time.
I want to take the claim seriously enough to actually check it, because the instinct behind it is not stupid. More confirmation feeling safer is a real intuition and it is sometimes right. The specific numbers being passed around, though, were never measured off anyone's trades.
First problem: nobody shows the sample
Here is the test I apply to any win-rate claim, mine included. Three questions:
- How many trades?
- Whose trades, on what instrument, over what period?
- What counted as a "confluence," decided before or after the trades were logged?
Go find the original source of the 70% number and try to answer those. I have looked. Every version traces back to another version. There is no study, no dataset, no journal export, no broker statement. The number is cited the way a proverb is cited.
None of that makes the claim false. It makes it unverified, and unverified should still feel uncomfortable. A number with no sample behind it tells you how someone feels about their setups, nothing more.
And here is the part that should really bother you: a claim like "three confluences produce 70%" is not even well-formed. Three which confluences? A trend filter, a session filter and a volume filter behave nothing like a trend filter, a momentum oscillator and a moving-average slope. One of those sets is measuring three different things. The other is measuring one thing three times.
That distinction is the whole ballgame.
Why correlated signals do not multiply
Let me show you where the 70% probably came from mathematically, because I think it came from an honest mistake rather than a lie.
Suppose you have a signal that is right 60% of the time. Not amazing, but real. Now suppose you have three of them, and they are independent — meaning each one carries information the other two do not have.
If you start from a coin flip and update on three independent signals that each point the same way, the math gets you to roughly 77%. That is the staircase. That is where "3 confluences = 70%+" lives. The arithmetic is fine. What breaks is the assumption underneath it, that the three signals are actually independent.
Think about what most people actually stack. Price above the 200 EMA. MACD above zero. Higher high on the daily. Those are three different indicators on three different panels of your chart, so they look like three pieces of evidence. They are three readouts of the same underlying fact: this market has been going up.
When the second signal is largely determined by the first, it adds almost nothing. In the extreme case where signal C is a mechanical consequence of signal A, seeing C after you have already seen A gives you exactly zero new information. Your probability stays at 60%. You just feel more confident about it, which is worse than useless, because now you size up.
Most real confluence stacks sit somewhere in the middle, partly overlapping and partly not. Which means the honest answer to "how much does the third condition add?" is: some amount between zero and a lot, and you cannot know it from theory. You have to measure it on your own trades. I wrote about how confluences interact rather than accumulate in why combinations matter more than individual signals, and that piece is the conceptual companion to this one.
The practical rule I use: before adding a condition, ask what it would take for this condition to disagree with the ones I already have. If you cannot picture that chart, the condition is decoration.
Second problem: the trades you no longer take
Say you accept all of the above and you go measure it yourself. Good. Now you hit the cost nobody prices in.
Imagine 200 logged trades. Condition A shows up in 120 of them. A plus B narrows it to 70, of which 40 are winners, so 57%. A plus B plus C leaves you with 22.
Twenty-two trades. Fifteen winners. That is 68%, and it will display in your journal as 68.2% because software loves a decimal.
Run the margin of error on that. A win rate of 68% over 22 trades has a confidence interval of roughly 48% to 88% — the confidence interval being the range the true win rate could plausibly sit in, given how little data you have. Forty points wide. The real number could be a coin flip. Your baseline over the full 200 trades, say 104 winners for 52%, has an interval of about 45% to 59%. Much narrower, because there's much more data behind it.
So the "improvement" from 52% to 68% is a comparison between a number you can half trust and a number you cannot trust at all. The two intervals overlap enormously. If that sounds like an accounting trick, it is not; it is just what small samples do, and I walked through the calculation by hand in your win rate has a margin of error so you can run it on your own filters.
This happens fast, too. Each condition you require cuts the surviving sample, and by the fourth tag most traders are looking at single digits without noticing, which is the failure mode I described in four tags and your sample is gone.
Most of us have done this. You filter down to the version of your strategy that looks brilliant, and the reason it looks brilliant is that you are now looking at eleven trades from one good month.
Third problem: the price of waiting
The cost that gets discussed least is the one that shows up in your P&L most directly.
Confluences arrive in sequence. The structural ones are there before you enter — trend, session, level, time of day. The confirmation ones arrive after price has already moved. A candle closing where you wanted it, a momentum cross, a lower-timeframe break of structure (price pushing past the last swing high or low). Waiting for the third confirmation usually means entering later and worse.
Concrete version. You are trading a level. Stop goes behind structure, 10 points away. Target is 30 points out. That is a 3R trade — R being one unit of risk, whatever you lose if the stop gets hit. Risk £100 to make £300 and you have made 3R. It lets you compare trades of wildly different sizes without converting everything to money first.
Now you wait for confirmation number three. Price moves 4 points in your favour before the signal prints, and you enter there. Your stop is still behind the same structure, so your risk is now 14 points, not 10. And your target is now 26 points away instead of 30.
26 divided by 14 is about 1.9R. You just gave up more than a third of your payoff to buy some confidence.
Now put both versions through expectancy, which is the only scoreboard that matters here. Expectancy is your average result per trade in R, winners and losers blended together — win rate and payoff in one number instead of two.
Before the numbers, be clear about what I am doing, since the whole complaint in this article is that nobody shows you the sample. I am bolting two separate illustrations together: the win rates come from the imaginary 200-trade journal above, the payoffs come from the level trade I just described. They are not the same set of trades. I am splicing them to show the shape of the trade-off, not to report a measurement.
- Two conditions: 57% win rate at 3R.
(0.57 × 3) − (0.43 × 1) = +1.28Rper trade. - Three conditions: 68% win rate at 1.9R.
(0.68 × 1.9) − (0.32 × 1) = +0.97Rper trade.
The win rate went up eleven points and the expectancy went down.
And notice we granted the confluence stack its best-case win rate, the shaky 68% from 22 trades. If the true number is 58%, the three-condition version earns (0.58 × 1.9) − (0.42 × 1) = +0.68R and it is not close.
I'm not claiming this outcome is universal. Some third conditions genuinely filter out a cluster of losers and are worth the worse entry. The point is that the trade-off exists and is rarely on the table when someone shows you the staircase. Win rate is the cheapest number to inflate and the least informative one to inflate.
So how many confluences should you use?
The honest answer is that "how many" is the wrong unit. Two conditions that measure genuinely different things will usually beat five that measure trend from five angles.
What I would actually do, in order:
Start from your unfiltered baseline. Every trade in your strategy, no conditions applied. That is the number every filter has to beat. If you skip this you have nothing to compare against and any filter will look good.
Add one condition at a time and record two things, not one. Win rate and average R. Also record how many trades survived. If your sample drops below about 30, write the result down as something to test later and keep trading the way you were.
Ask whether the condition earns its cost. A condition that lifts win rate 6 points while cutting your average winner from 2.4R to 1.7R has made you worse off. That single comparison kills more "high-probability" setups than anything else I know.
Check whether one condition is doing all the work. Often you will find that a three-condition stack performs the same as the one condition inside it that actually matters, which is the exercise in finding which condition carries your edge. Then you can drop the other two and get your sample size back.
Validate on trades you have not looked at yet. Any combination discovered by hunting through your own history is guilty until proven otherwise. Splitting your data chronologically and testing the rule on the later half is the minimum standard, and there is a fuller walkthrough of the process in testing condition combinations properly.
That fifth step is the one people skip, and skipping it is how a lucky run gets promoted to a rule. Nobody can promise you which conditions will keep working. All you can do is stop fooling yourself about which ones have worked so far.
Measuring whether an added condition actually improves expectancy, instead of assuming more confirmation is better, is precisely what EdgeFlow's combination testing reports. It tags each trade across setup, execution, environment and management, then shows you a likely range rather than a single number, and tells you plainly when a filter has left too little data to say anything at all.
The staircase graphic will keep circulating, because it flatters an instinct we all have. Your own journal can answer the question directly, though, and it will answer it about your trades rather than someone's carousel.
Start with the baseline. Add one thing. See what it actually costs.