Loading…
Reading the draws and working out what chance predicts.
Loading…
Reading the draws and working out what chance predicts.
Every draw has a face
A draw is six numbers, and it is also a face: how much it adds to, how it splits odd and even, how tightly it clumps, how wide its emptiest stretch runs. Here is the latest one, read feature by feature against every ticket Big Sky Bonus could ever produce.
Typicality 42/100 — a description of this face, not a forecast.
Big Sky Bonus4/31 + 1/16 era · since Oct 2024734 draws
Point at the galaxy to read a draw.
Fingerprint your own ticket
Pick 4 numbers.
The last pitch in the book
“Look at the winning draws: 40% of them have exactly 2 odd numbers. Random tickets waste their money — balance your 4 numbers 2 odd and 2 even, and play the way winners play.”
Every number in that quote is true. Press the button.
The two layers underneath
Every statistic on this page — the sum, the odd count, the widest gap — is a map from a combination to a number. Many combinations share each value, so each map is many-to-one, and each one flattens the 31,465-ticket space onto a single axis.
The Atlas is what you get by pushing the uniform distribution over combinations through those maps: for every feature value, how many tickets carry it. And because a winning draw is a uniform pick from exactly that space, the features of winners must follow the Atlas. Not approximately, not usually — by construction.
That is the whole theorem of this feature. Everything below is a consequence.
A pool of 31 numbers holds 16 odd positions and 15 even ones. Choosing odd means choosing from the odds and from the evens:
That is the hypergeometric distribution, and the low/high count works the same way with the pool split at its midpoint.
Note the asymmetry when the pool is odd-sized. A 1–49 pool has 25 odd numbers and 24 even, so a perfectly "balanced" 3-and-3 split is very slightly more likely than its mirror — a detail no balance-system pamphlet has ever mentioned, because it does not help.
Under fair draws the points can only ever fill the nebula. They cannot form a band, drift toward a corner over time, or leave a hole, because each draw is an independent uniform pick and independence has no memory to drift with.
It is worth being precise about what would show up if something were wrong. A generator constrained to a subset of outcomes — the Tipton case is the documented example — does not produce impossible draws. It produces draws that avoid regions the atlas says should be populated, or crowd regions it says should be sparse. That is a real, testable signature, and it is what a randomness audit looks for. On these eras there is nothing to find, which is the honest and slightly boring result.
This is the most reasonable question a modern reader can ask, and it has a clean answer.
A learner needs structure: some function from past to future that beats guessing. Draws here are independent and identically distributed, uniform over 31,465 outcomes. Under that model the conditional distribution of the next draw given the entire history is the uniform distribution — the optimal predictor is the one that ignores the history completely, and no amount of capacity improves on it.
So a model trained on draw history can only do one of two things: reproduce uniformity, or overfit. There is a small published genre of neural-network lottery-prediction papers, and their reported accuracies are what overfitting noise looks like when it is scored on the data it memorised. The tooling is genuine and the data is honest; the target is what does not exist.
Feature space is where data science starts. For a fair lottery, it is also where it honestly ends.
People read faces into things. A row of six numbers is no exception: 1-2-3-4-5-6 looks like it can't win, and something scattered across the middle looks like it might. Ask anyone which of the two they would rather hold and you will get an answer, confidently, immediately.
Both are one ticket. Both have the same chance. The feeling that one has a better face is the representativeness heuristic — the mind judging a specific outcome by how well it resembles the process that produced it, rather than by its probability. A scattered ticket looks more like "randomness", so it feels more likely to be produced by randomness.
What the Galaxy adds to that familiar argument is scale. It is not one feature being defended; it is a full portrait, six dimensions at once, every draw the era ever produced — and every face on it belongs to the same family. There is no region of the nebula where the winners cluster. There is only the nebula.
Out of that instinct grew a small industry. Books, subscription sites and app listings will advise you to split your numbers evenly between odd and even, high and low, and will support the advice with real statistics from real draws: most winning tickets are indeed roughly balanced.
The arithmetic behind the pitch is the same one this playground has now taken apart three times, in three costumes. With sums it was "winning totals cluster in the middle". With deltas it was "winning gaps are small". Here it is "winning numbers are balanced". Each is a true fact about winners, and each is equally a true fact about every losing ticket ever printed, for the same reason: there are more ways to build a ticket that way.
The advice does have one measurable effect. It moves you toward the crowded part of the ticket space, where more people are standing, and where a jackpot is split more ways. It is the rare kind of strategy that is not merely useless but mildly counterproductive.
Every couple of years the story comes round again: a neural network has learned to predict the lottery. There are papers, apps, videos, and occasionally a confident founder.
It is worth telling kindly, because nothing about it is stupid. The tools are real and powerful. The data is clean, complete, and free. The problem looks exactly like the ones machine learning has genuinely conquered — a sequence, with structure to find.
The difference is that here there is nothing to find, and no model can be told that in advance. Feed a flexible learner pure noise and it will not report failure; it will report a pattern, fitted beautifully to the past and worthless on the future. The Galaxy is what their training data actually looks like: a nebula with no shore, no current, and no direction.
You cannot out-model a coin.
The draws have faces. The faces have statistics — real ones, exactly computable, sometimes surprising, and this whole page is built from them.
The statistics also have limits. Knowing where those limits fall is the only genuine expertise the lottery permits, and it is worth more than any system, because it is the one thing on offer that is actually true.