Do Patterns Exist in the Lottery and How to Look for Them
The suspicion that a lottery hides a secret pattern occurs to everyone who looks seriously at the draw archive. The number 17 came up 14 times in September, while 3 did not appear at all. Three odd numbers in a row showed up 6 times over the year. The pair 12-23 appears especially often. It feels like an algorithm is hiding somewhere — just carefully disguised. The short answer: in a fair lottery there is no algorithm, but the noise of randomness genuinely produces pictures that look a lot like a pattern. The site has five tools that test different types of possible patterns one by one, and they all reach the same conclusion — with varying degrees of certainty.
The Main Answer and Why Noise Looks Like a Pattern
In a fair lottery, the outcome of the next draw is independent of all previous ones. This is the mathematical definition of randomness, not a philosophical claim. The drum has no memory of history: every combination has a chance equal to the reciprocal of the number of combinations, and that chance does not change based on what came up yesterday.
But randomness produces shapes. If you flip a coin 1000 times, somewhere in the middle there will inevitably be a run of 7 heads in a row — and this is an absolutely normal phenomenon, not "rigging". The reason: in a long sequence all possible short shapes occur at their expected frequency, and with a large sample rare events become almost inevitable. The brain is good at spotting these shapes and bad at understanding how expected they actually are. This effect is called apophenia — the tendency to see patterns where none exist.
Hence the practical distinction: you can always describe a pattern after the fact (any sequence is describable), but you cannot predict it in advance unless it has a causal structure. And the test of "does this pattern have a causal structure" is exactly the problem that statistics solves.
How to Tell Noise from Signal: P-value and Confidence Interval
Every statistical test for randomness works the same way. First, a null hypothesis is formulated: "the data are random". Then it is calculated how likely the observed result is under that assumption. This probability is called the p-value. If the p-value is very small (the standard threshold is below 0.05), it means the data fit the null hypothesis poorly, and the hypothesis can be rejected.
In practice, the p-value must be interpreted carefully. At a threshold of 0.05, every twentieth fair test will, by chance, produce a false "significant" result. If you run 1000 such tests on different combinations of numbers, in 50 of them a "pattern" will appear on its own — not because it exists, but because you looked 1000 times. This is called the multiple comparisons problem, and it is the main source of false conclusions about patterns in the lottery.
So a real test is not "found something with p < 0.05", but "found something with a p-value corrected for the number of tests". And here is where it gets interesting: all five tools on the site are built precisely with this correction in mind. For more on the principles of multiple testing, see the article on 20 methods of analysis.
Autocorrelation: Does a Draw Depend on the Previous One
The most direct test for the existence of a "number-drawing algorithm". The autocorrelation section computes the correlation between adjacent draws: if today's draw has a causal link to yesterday's, it will show up as a persistent deviation of the autocorrelation function from zero.
What the real data show. In MillionDAY (5 of 55), with thousands of draws, the autocorrelation fluctuates around zero within the confidence interval — it is noisy, as a random process should be. No lag (delay) gives a persistent signal. If a "drawing algorithm" like "after 17, 23 comes up more often" existed, we would see it here as a peak at lag 1. There is no such peak.
In Keno, where draws happen every few minutes and the archive is enormous, the autocorrelation also fluctuates around zero, but the confidence interval is narrower — we could detect weak effects. And we do not. This is one of the most convincing results, because with such a volume of data even a tiny pattern would be visible.
The Runs Test: Randomness of Alternation
Autocorrelation looks at the numbers themselves, while the runs test looks at the shapes of their distribution. For example: how long do runs of even (or odd) numbers get in the archive? How long do runs of "high" and "low" get? In fair randomness there is a theoretical distribution of run lengths — and a deviation from it indicates that independence is broken somewhere.
The result across the main number lotteries is the same as for autocorrelation: the deviations stay within the confidence interval. Runs of 5-6 numbers of the same parity occur exactly as many times as randomness would predict. The same goes for "runs of adjacent numbers" and "runs from the same range".
The runs test is especially good at spotting two things. First, artificial regularity: if a drum were programmed to "avoid" repeats, long runs would occur less often than expected. Second, an equipment artifact: a physical defect in the drum would create "clustering" that boosts long runs beyond expectation. Neither of these is visible in our archives.
Benford's Law: Does It Work for Lottery Numbers
The most exotic of the five methods. Benford's law states: in "natural" numerical data the first digit is distributed unevenly. The digit 1 appears in about 30% of cases, 2 in 17%, 9 in 5%. The law works beautifully for incomes, river lengths, stock prices — anywhere numbers span several orders of magnitude.
For lottery numbers from 1 to 55 or 1 to 90, the law strictly speaking should not work. The range is too short, and the digits "1-9" do not span an order of magnitude. But a graphical comparison of the actual distribution with the theoretical one is still useful — it shows how statistical laws behave on limited ranges and where they break down.
In MillionDAY the first digit of a number is distributed almost evenly (1, 2, 3, 4, 5 each about 18%, with the rest split across 6-9). This is exactly what "fair randomness within a limited range" predicts. There is no magic — but it is interesting to see how a general rule adapts to its context.
Markov Chains: Transition from Number to Number
The most complex of the tests mathematically, the most intuitive in concept. Markov chains build a transition matrix: if the number 12 came up in a draw, with what probability will 25 come up in the next draw? For each pair (i, j) out of 55×55 = 3025 pairs, this probability is computed.
In a fair lottery all 3025 probabilities should equal the theoretical value (roughly 5/55 ≈ 0.091 for MillionDAY — the probability that a specific number appears in a draw at all, independent of the past). The Markov analysis builds this distribution and checks whether it is uniform.
The result is predictably boring: the matrix is uniform, all probabilities are close to the theoretical value, and the deviations stay within the confidence interval. But the method itself is useful not for "finding patterns" but for visualizing what randomness looks like at large scale. When you look at a 55×55 heatmap with a uniform background, the intuitive feeling that "an algorithm is hidden somewhere here" fades. The archive really does look random.
The five methods above cover five different types of possible patterns. If even one of them showed a persistent signal with a p-value corrected for multiple comparisons, that would be a serious claim to a discovery. But across all large archives all five return "randomness", which boils down to a single table:
Test | Which pattern it looks for | What it would show if present | Actual result |
|---|---|---|---|
Dependence of a draw on previous ones (lags 1, 2, 3 and beyond) | A persistent peak at one of the lags | Fluctuations within the confidence interval | |
Regularity of runs (even/odd, high/low) | Runs that are too even or too long | Runs distributed as in fair randomness | |
Structural distortions in the distribution of first digits | A strong deviation from the Benford model | A uniform distribution, as expected within a limited range | |
Conformity of frequencies to a uniform distribution | A large chi-squared value and p < 0.05 after corrections | Frequencies converge to uniform on large archives | |
Dependencies in pairwise transitions (i → j) | A non-uniform transition matrix | The matrix is uniform, all transitions equally likely |
Five independent methods, five different "slices" of the idea of a pattern, one and the same conclusion. This is not a coincidence of matching results — it is a property of the data.
The Honest Conclusion: Why the Lottery Resists "Hacking"
The logic is simple and, oddly enough, economic. A lottery operator earns on the difference between ticket sales and prize payouts. If an extractable pattern existed in the lottery, sooner or later it would be found — by scientists, analysts, algorithms. A few successful players would get rich, and the lottery's expected value for the operator would go negative. The lottery would quickly shut down because it had become unprofitable.
The fact that major lotteries run for decades and stay consistently profitable is indirect but powerful evidence of the absence of exploitable patterns. Drums are regularly inspected and replaced, RNG algorithms in digital lotteries are certified to standards, and archives are open to independent analysis. All of this works precisely because "hacking" is impossible.
But it does not follow from this that statistics are useless to a player. They are useful in a different way: they help filter the choice. If there are no patterns, then any choice of numbers is equivalent to the rest in jackpot odds, but some approaches are more systematic, discipline the budget, and reduce the chance of splitting a prize with the crowd. For more on sensible approaches, see the article on playing strategies.
And one last thing. If somewhere online you see an "algorithm for predicting the lottery", pay attention to how it is tested. An honest test predicts future draws, not a fit to already-played history. Finding an algorithm on the archive and then applying it are completely different tasks. On already-played data you can "explain" any sequence, even a random one — this is called overfitting and is common among many "lottery AIs", as discussed in detail in the article on the neural network for the lottery.
The Bottom Line
There are no patterns in a fair lottery. This follows from the definition of randomness and is confirmed by all five tests on the archive.
The noise of randomness produces shapes that visually resemble patterns. Apophenia is the brain's tendency to notice and exaggerate them.
The p-value must be corrected for the number of tests. With 1000 checks, 50 "significant" ones will be found purely by chance if no correction is applied.
Five tools (autocorrelation, runs test, Benford, Pearson, Markov chains) test different kinds of patterns and consistently return "randomness" on all of the site's number archives.
Lotteries are economically resilient: if a pattern existed, the operator would go bankrupt. This works as indirect evidence.
Statistics do not predict numbers, but they help discipline the choice and weed out rituals masquerading as strategies.
Testing an algorithm on the archive is always a fit. The only honest test is predicting the future, and in that format every "algorithm" performs like random guessing.
