Nine questions to ask any trading-statistics product — including this one
We could have written a comparison table. We had the research for it: this category's pricing and catalogue sizes are public. But those are marketing pages, and "they never show a sample size" is a claim about a product we have not logged into — unfalsifiable by the reader and uncorrectable by the subject. So here is the useful version instead: the questions that decide whether a percentage means anything, our own answer to each, and a link to where you can check it.
The nine questions
Does every percentage carry its sample size?
A 70% win rate on 25 occurrences and a 70% win rate on 25,000 are the same number and different facts. Without n you cannot tell them apart, and the smaller one is far easier to find.
Every figure on this site ships with its n, and windows under 300 samples are not published at all. The dashboard prints a 95% interval beside the percentage, so you can see how wide it is.
Check it — the coverage table, market by market →What was the number measured against?
This is the one that decides everything. "Weekend gaps fill 88% of the time" is true — and an ordinary midweek move of the same size fills 88% too, so the statistic is about markets in general rather than about weekends. Almost every claim in this field dies against a control, and almost nobody prints one.
Every comparison here runs against a matched control: a size-matched move, a permuted grid, a direction-randomised day, an era the selection never saw. 26 of our 62 tested claims do not survive theirs, and they are on the site with the rest.
Check it — all 62 of them, filterable →How much history — and can you change it with a dial?
A lookback selector feels like a feature and behaves like a search. Shorten the window and the sample under every bucket shrinks with it, so patterns appear that the full history does not support.
Twenty-three years, fixed — and we measured what a dial would do to our own grid, at the settings this category's dials actually reach. At a five-year lookback the median weekday-hour window holds 217 candles and 0% of them clear our 300-sample floor. Drop the floor — which is what the category ships, since none of it prints a sample size — and the five-year dial would put 102 windows on a screen, 33.3% of them unsupported by the full history. And it is not recency. An equally long block of years from earlier in the same history, sharing not one year with the recent one, shows 78 with 41.2% unsupported — as many, and as unsupported. The dial sells a recent market and delivers a short sample. At ten years, where it finally clears our floor, it would call 127 windows significant, 38 of them unsupported, while hiding 91 that are. That is why there is no dial.
Check it — every threshold we could have chosen, and where ours sits →Does it publish the tests that failed?
A catalogue where everything works is a catalogue that stopped looking. The failures are what tell you the method is capable of returning "no".
Most of this site is negative. Of 62 tested claims, 26 do not survive and 23 only half do. The day of the month, the opening range, the gap fill, yesterday's high, the hot streak and the opening candle all failed here.
Check it — filter it by "does not" →Does it tell you when a published number changed?
These figures are rebuilt as new data arrives. If a window you screenshotted last month is no longer flagged, you are owed the reason in public rather than a quietly re-rendered page.
A changelog with 40 corrections in it — entries that moved a number somebody may already have read — plus an Atom feed, so you can subscribe to us being wrong.
Check it — every correction, dated →Can you download the data and redo the arithmetic?
A number you cannot recompute is a number you have to trust. The whole argument of this site is that you should not have to.
Every aggregate we publish is a plain JSON file, CORS-open, free for any use with attribution, and each one carries its own source and licence inside it. No key, no tier, no rate limit.
Check it — the files, and the contract they ship under →Does it say where it does NOT work?
A tool that returns an answer for every symbol you type is telling you about its interface, not about the market.
We grade every market and mark the ones that fail. 13 can demonstrate hour-of-week structure and 6 cannot — those are labelled in the picker, in the embeds, in the articles and in the extension. One market cannot produce a single publishable bucket and we say so on its own page.
Check it — the minimum lean we could even detect, per market →Does it charge the cost of trading before showing you the edge?
A lean smaller than your spread is a loss with a good chart. Gross numbers are the easiest way to make a weak result look strong.
Nothing net-negative is published, every rule states the spread it stops paying at, and you can set your own broker's cost and watch the list shrink. At three times our assumption the screener goes from 120 rows to 3.
Check it — what cost each edge dies at →Does it publish what happened AFTER it went live?
Every advertised hit rate in this field comes from a backtest, and a backtest is the number a vendor has the most control over. What the product actually produced once it started running is the one figure nobody can tune in advance — and the one nobody is obliged to show you.
Our alerts have run since 2026-08-07. The 12 windows the rail pushes have come round 60 times and closed the way the rule said 41 — 68.3%. Read as 60 independent tries that means 55.8%–78.7% — but they are 60 occurrences of just 12 windows, one of which supplies 18 of them, so resampled at the level we actually picked it is 42.6%–85.7%, which does not rule out a coin. The 21 windows that clear the same bar and are deliberately not pushed are scored beside it as the control, at 58.1% — and those two ranges still overlap, so the gap is not yet evidence our filter earns its place. It is free to read whether or not you ever buy the alerts.
Check it — the forward record, and the control beside it →What it costs, and what that buys
Ours is $79 a year or $199 once, and the free tier is not a demo: it carries the full 23-year history on the four major pairs, with the same sample sizes, the same intervals and the same significance flags as the paid one. What Pro adds is more markets, the strategy board, the five-minute view and the alerts.
What we are worse at
The catalogue is smaller. We publish 62 tested claims. The larger platforms in this field publish well over a hundred reports, and if you want breadth of setups rather than depth on a few, they have it and we do not.
Nineteen markets, not thousands. FX majors and minors, gold and silver, five indices and two crypto. No individual stocks, no futures contracts, no ETFs.
Hourly, not tick. We hold five-minute data on four pairs and nothing finer, so anything that lives inside a five-minute bar we cannot answer — and we measured what happens when you cut the clock finer: the buckets shrink and the spread does not.
Direction is the weak half of our own product. Our out-of-sample test values ordinary-hour direction at +0.56 points, which is less than a round trip. The half that survives is when a market moves and how far. We would rather you knew that before paying than after.
No signals, no automation, no broker connection. Nothing here places a trade or tells you to. If that is what you are shopping for, this is the wrong product.
The whole free tier is 23 years of history on the four majors, with the sample size and the interval on every figure. Nothing below is behind a form.
Open the dashboard →