Loading…
Reading the draws and working out what chance predicts.
Loading…
Reading the draws and working out what chance predicts.
Methodology
Every figure on this site comes out of the same few rules. They are set out here in full, in the order the data passes through them: from a draw reported by several sources to a chart that says whether a pattern is real. Each tool’s own page carries the mathematics behind its particular test.
A draw arrives as raw records, one per source. Sources are ranked by trust — official files and APIs first, then a commercial feed, then community mirrors — but rank does not decide a draw. Records are grouped by the exact numbers they report, and the reading most sources give wins; only a tie breaks toward the more trusted source. That rule exists because of Mega Millions on 10 May 2022, when an official portal reported Mega Ball 6 while two other sources reported 9.
Every settled draw carries one of four tags:
A record whose numbers do not fit the game’s rules — the wrong count of balls, a ball above the pool, a duplicate — is refused, never repaired into something plausible; it is counted as a format violation on the data page. Two patterns the feeds produce are recognised and set aside rather than promoted: the same result filed under two sub-game labels on one night, and a draw posted again under the next date. Every night an integrity scan checks the whole record for the ways it has been found wrong before.
When a lottery changes its pool — UK Lotto went from 49 balls to 59 in 2015, Powerball from 59 to 69 the same year — every frequency changes with it, and a count that mixes the two eras means nothing. So no statistic on this site crosses a rule change. Each game’s history is divided into eras, a tool reads one era at a time, and the era it is reading is named on the page.
For the flagship games the era boundaries were verified by hand against the lotteries’ own records. For the rest they are derived from the draws: a ball above everything seen before, once an era has enough draws to have surely shown its top ball, proves the pool grew. A pool that shrank cannot be told from a top ball that has not come up lately, so a derived era keeps the larger pool — which never rejects a real draw but can overstate the pool, and is why such games carry a “pool inferred” caveat wherever their scope is named.
No count is shown without its expectation. In a game drawing k numbers from N over D draws, each number is expected D·k/N times, and the gap between that and what happened is measured in standard errors of the binomial count — the σ that appears beside every figure. Up to 1.5σ is ordinary; between 1.5σ and 2.5σ is notable, the kind of gap one cell in twenty shows; beyond 2.5σ is extreme. The colour ramp runs grey to amber to violet and never green to red, because a number drawn more often than expected is not a lucky number.
A surface that shows sixty-nine numbers at once is sixty-nine simultaneous tests, and about one in a hundred will clear a 1% threshold by luck alone. So the threshold is divided by the number of tests on the surface (the Bonferroni correction), and the badge on every figure says how many tests it stands among. A finding that survives that correction is called out as such; one that only looks unusual on its own is labelled as exactly that.
Waits between appearances follow the geometric distribution, with mean N/k draws and no memory: having waited fifty draws says nothing about the fifty-first. A wait still running is a lower bound on a gap that has not finished, and is never averaged in with completed waits.
A rule change resets the clock, and a chart drawn from 144 draws is noise wearing the costume of a finding. Before a surface renders, the site asks how often chance expects its smallest cell to have been seen. A number-frequency or waiting-time chart needs each number expected about 30 times to render without caveat and at least 10 to render at all; a pair chart, where there are thousands of cells, needs each pair expected 5 times, and 2 to render; a distribution needs 200 draws, and 50 to render. Below the floor the page says what is missing and how many draws would fix it, rather than showing a chart it cannot stand behind.
Lotteries do not publish which numbers people pick, so this used to be borrowed from the research literature. It is now measured from the games themselves. A prize table says how many people won at each tier; winners against that tier’s probability say how heavily the drawn numbers were held, and fitting that across a whole era recovers how much more than average each number is played. Every coefficient, its standard error and the draws behind it are on the edge page. On Powerball the fit puts a day of the month about 14% above an average number and a number above 31 about 7% below — the birthday effect, measured rather than asserted.
What it assumes: that the habit is stable across an era, that it is the same for every draw of that era, and that a tier’s winners are dominated by people choosing their own numbers. Where the record cannot support a fit the site says so and falls back to the literature model; of the ten fits held, five are marked reliable and the rest are shown as unreliable rather than quietly used.
The caveat that matters: a quick pick is drawn by the terminal, not chosen, and quick picks are a large share of tickets. The fit therefore measures the crowd it can see — the chosen lines — diluted by every random one beside them. Read it as the direction and rough size of a habit, never as the number of people holding any particular line.
Several operators name the drum and the set of balls used for each draw, and the site keeps both. A machine is not a rule era: the same rules can be drawn on four machines, and counting them together hides a drum that behaves differently from its siblings. The Machine Room reads each machine against the others under one set of rules, which is the only comparison that means anything. Where a feed sends no machine, nothing is inferred.
A prize row arrives as prose — “Match 4 plus Bonus”, “rang 3”, “Division 2” — and means nothing until it is keyed to an outcome. Every row in the archive carries a key saying which numbers and which bonus balls it pays for, or a recorded reason why it could not be read. The coverage, and the games that fall short of it, are on the data page. This is what implied sales rest on: winners at a tier whose players matched none of the main numbers, divided by that tier’s probability, estimate the lines in play without any model of how people choose. That figure is the denominator under the return ledger, the add-on take-up and the crowd fit, which is why its coverage is published rather than assumed.
Prize amounts and jackpots are shown as the sources reported them, in the lottery’s currency, and are not financial advice. The operator’s published result is final on any question of a prize.