Somebody posts "prior day high, 62% chance it gets hit." Sounds strong. You have no idea whether it is, because you weren't told the one thing that decides it: how often that happens anyway.
If price hits the prior day high on 57% of all days, then 62% on your setup is worth five points. Mildly interesting. If it happens on 80% of all days, that same 62% means your setup made it less likely, and somebody just sold you a headwind as an edge.
Same 62% twice. One is a small tailwind, the other is a real headwind. We print the base rate next to every number so you never have to take the headline on trust.
Delta is the gap between the two: your number minus the base rate. It's the only bit your conditions actually caused.
A +18 on a setup means those conditions moved the odds eighteen points past what a random day would give you. A −18 means they moved it the other way, which is just as useful once you stop reading it as failure. Knowing a thing is unlikely today is worth as much as knowing it's likely.
We sort by the size of the delta, ignoring the sign. The biggest movers rise to the top whether they push for or against.
Stack enough conditions and you can get any number you like. Twelve filters on twenty years of data might leave you four matching days. Three of them went your way. The screen says 75%.
It means nothing. Four days is four days, and you'd have found something that looked just as convincing if you'd flipped a coin four times.
When you click TODAY, we apply your conditions in order and stop before the sample breaks. Then we tell you what we dropped. Twelve of twenty applied, eight dropped to keep the sample above 60. Most tools would apply all twenty, land on one matching day, and print a confident 100%.
The 95% range next to each number is the honest width of the answer. Read it as: given this sample, the true rate is somewhere in here.
Both say 71%. The first is pinned down between 64 and 76 and you can plan around it. The second could be anywhere from a coin flip to a near-lock, and you'd be guessing which. Wide interval, small sample, same story told twice.
A pattern that printed beautifully from 2012 to 2017 and has done nothing since will still look excellent averaged over twenty years. The average hides the death.
So we cut the history into five eras and check each one separately. The five squares on every row are that test. Green means the effect held in that era, amber means it didn't, grey means too few days there to judge.
Five green squares beats a bigger number with two. A pattern that survived 2008, 2020 and the last two years has been tested by conditions you can't reproduce on purpose.
Test one idea and a big result means something. Test four hundred and keep the biggest, and a big result means nothing. You were always going to find one. That is not bad luck, it is arithmetic. The more places you look, the larger the best thing you find, even in data with no pattern in it at all.
This is the single most common way traders fool themselves, and it usually doesn’t feel like searching. Flicking through setups until one looks good is the search. You just didn’t count the ones you rejected.
So every search here is run again on a shuffled copy of itself, hundreds of times, and we record the biggest thing luck managed. That becomes the bar, and it is the bar for the whole search, not for one row. Look at more things and it gets harder to pass, exactly as it should.
The oldest promise in trading is that you can fix a losing strategy with better risk-reward. Aim further, risk less, and let the maths do the rest. It does not work, and you can watch it not work.
Widen your target and your win rate falls. Tighten your stop and it falls again. We used to say the exchange rate between the two was almost exactly fair. Measured properly across 384 settings it is not: it runs against you, and the tighter the stop the worse it gets.
A reward-to-risk under 0.75 comes in 1.9 points ahead of what the arithmetic promises. Over 2.2 it comes in 4.0 points behind. Six markets, same sign, no exceptions. So a 2:1 that the calculator says needs 33% actually pays about 29%, and a setup that looked break-even on paper was never break-even.
Chasing a bigger target raises the bar faster than it raises the hit rate. What geometry decides is the shape of your results: many small wins or few large ones. What it cannot decide is whether you have an edge. Only a reason to be in the trade does that.
We checked whether any of our 19 setups changes this. Across 568 combinations, 1% cleared two standard deviations, which is below what a search that size turns up by chance. Conditioning does not move it. Price your own geometry against fair value and see where yours lands.
Ask how often an hour’s high forms in its first fifteen minutes. A quarter of the hour, so a quarter of the highs, or 25%. Measure it and you get 37%. Half again as often as it should be. Looks like a real pattern about how the hour opens.
It mostly isn’t. The moment a random wander makes its highest point is not spread evenly across a window. It bunches at the start and the end and is rarest in the middle. Pure mathematics, no market involved. A coin-flip path already puts about a third of its extremes in the first quarter.
The effect is real. It holds in all six markets and all five years. It is just a third the size the obvious comparison suggests. When someone shows you a number without telling you what it should have been, that is the question to ask.
Every probability here is the chance price touches a level. Not the chance you make money going there.
Price can tag your target after stopping you out first. An 80% touch rate with a stop in the wrong place is a losing strategy with a beautiful hit rate, and we've tested exactly that: a signal we measured at 79% accuracy returned roughly nothing once you accounted for entry price and stop distance. The accuracy was real. The money wasn't.
So on the Stops page we don’t ask which level price touched. We ask which one it reached first, bar by bar. A session that stopped you out at 10am and printed your target at 2pm is a loss, and it is counted as one. Almost nowhere else does that, which is why almost every published hit rate is better than the trade that produced it.