Published · · Last reviewed · · 11 min read
English · 한국어
What numbers do people pick?
The clearest result: people avoid the numbers that were just drawn. When more of a round's winning numbers repeat those of the previous draw, there are 1.3% fewer fourth-prize winners per repeated number (95% interval: 0.8% to 1.7% fewer), and 1.9% fewer third-prize winners. This fits the gambler's fallacy showing up in real purchases: the belief that numbers which just came up are "used up". The data show the pattern, not the reason; see the gambler's fallacy for why the belief is wrong: every draw is independent.
Every Lotto 6/45 combination has exactly the same chance of being drawn, 1 in 8,145,060 for the first prize. What differs is how many other people hold the same numbers. When many tickets carry a combination, the prize that combination wins is split between more people. So the question this page asks is not which numbers will come up, but which numbers people choose, and whether the official data can show it.
The idea is simple. If tickets were spread evenly over all combinations, a round's sales would tell us almost exactly how many people win each prize. The official figures do not behave that way: the number of fourth-prize winners varies around that expectation with a standard deviation of 7.7% from round to round, while pure chance would give only 0.33%. Something systematic is going on, and some of it can be traced to the numbers that were drawn.
This is research about human habits, not a way to forecast a draw. Nothing below changes the chance that any ticket wins.
In short
- Numbers repeated from the previous draw: 1.3% fewer fourth-prize winners (per extra number).
- Numbers up to 12: 4.1% more fourth-prize winners (per extra number).
- Spread of the numbers: 2.8% fewer fourth-prize winners (for every 10 numbers of extra spread).
- Numbers that share a last digit: 1.3% more fourth-prize winners (per extra number).
- The common story that people lean on birthdays (numbers up to 31) is not confirmed.
The data
For every round the official data give the total sales, the numbers drawn and the number of winners and prize money for each of the five prizes. We used rounds 88 to 1,250: 1,163 rounds, about 85 billion games sold. Rounds before 88 were left out because a game cost more then and the rules differed (87 rounds). No other round was excluded. Where the official data also split first-prize winners into automatic, manual and semi-automatic tickets (from round 262 on), the split is kept in the public dataset, but the analysis below uses the winner counts of the prizes themselves.
How the comparison works
A game costs KRW 1,000, so sales divided by 1,000 gives the number of games sold. Out of the 8,145,060 possible tickets, 1 win the first prize, 6 the second, 228 the third, 11,115 the fourth and 182,780 the fifth. Multiplying the games sold by those shares gives the expected winners of each prize if every combination were equally popular. The popularity index of a round and prize is the actual number of winners divided by that expectation: 1.10 means 10% more winners than an even spread would give.
The third and fourth prizes are the main measures because they have enough winners per round (thousands for the third prize, tens of thousands for the fourth) for the index to be precise. The fifth prize, with millions of winners, is used as a robustness check. A caution that applies to all of them: a fourth-prize ticket matches only four of the six drawn numbers, so the index reflects how popular parts of the winning combination are, not the whole combination.
We describe each winning combination with 12 simple features, such as how many of its numbers are 12 or below or how widely the numbers are spread, and ask a statistical model which features go with more or fewer winners than expected. To avoid being fooled by chance, the model is built on the earlier rounds and checked on the later rounds it has not seen; only effects that hold up on both are reported. The technical details are below.
Detailed method
We describe each winning combination with 12 simple features: how many of its numbers are 31 or below, 12 or below, neighbouring pairs, whether it contains an arithmetic progression, how many share a last digit, how many are multiples of 7, the largest number in one row or one column of the Korean play slip (seven numbers per row), how many pairs touch on the slip, whether three sit on one straight line of the slip, how many numbers repeat from the previous draw, and how widely the numbers are spread. A negative-binomial regression, which allows for more scatter than pure chance, relates the number of winners to these features. The expected winners enter as a fixed offset, and the games sold and a time trend are included so that bigger rounds and changing habits over the years are not mistaken for number effects.
Because twelve features were tried, some will look important by luck. Two rules guard against that. First, the model is built on the earlier 697 rounds (88 to 784) and checked on the later 466 rounds (785 to 1,250) that it has not seen. Second, a feature counts only if it stands out on the earlier part after correcting for the twelve comparisons (false discovery rate 0.10), and then keeps its direction on the later part with a Holm-corrected p-value below 0.05. Every interval below is a 95% interval that allows for the extra scatter between rounds.
What held up
4 of the 12 features passed both tests for the third and the fourth prize, with the same direction: numbers up to 12, spread of the numbers, numbers that share a last digit, numbers repeated from the previous draw. The chart shows the fourth-prize effects of all features; filled dots passed.
Show the numbers
| Feature | % | 95% interval | Status |
|---|---|---|---|
| numbers 12 or below | 4.1 | 3.6 – 4.7 | Confirmed on the second part |
| numbers sharing a last digit with another | 1.3 | 0.9 – 1.8 | Confirmed on the second part |
| contains a 3-term arithmetic progression | 0.9 | 0.2 – 1.6 | Selected, not confirmed |
| most numbers in one slip column | 0.8 | -0.1 – 1.7 | Not selected |
| numbers 31 or below | 0.5 | 0.1 – 0.9 | Selected, not confirmed |
| pairs touching on the slip (8-neighbour) | 0.3 | -0.1 – 0.6 | Not selected |
| three on one straight line of the slip | -0.1 | -0.9 – 0.8 | Not selected |
| most numbers in one slip row | -0.7 | -1.6 – 0.2 | Not selected |
| neighbouring pairs (n, n+1) | -0.9 | -1.5 – -0.3 | Not selected |
| multiples of 7 | -1.0 | -1.5 – -0.5 | Selected, not confirmed |
| numbers repeated from the previous draw | -1.3 | -1.7 – -0.8 | Confirmed on the second part |
| spread (largest minus smallest, per 10) | -2.8 | -3.4 – -2.1 | Confirmed on the second part |
Numbers up to 12. Per extra number, fourth prizes had 4.1% more winners than expected (95% interval: 3.6% to 4.7% more), and third prizes had 5.3% more winners (95% interval: 4.3% to 6.3% more). On the held-back rounds both prizes moved in the same direction.
Spread of the numbers. For every 10 numbers of extra spread, fourth prizes had 2.8% fewer winners than expected (95% interval: 2.1% to 3.4% fewer), and third prizes had 4.2% fewer winners (95% interval: 3.1% to 5.4% fewer). On the held-back rounds both prizes moved in the same direction.
Numbers that share a last digit. Per extra number, fourth prizes had 1.3% more winners than expected (95% interval: 0.9% to 1.8% more), and third prizes had 2.1% more winners (95% interval: 1.3% to 3.0% more). On the held-back rounds both prizes moved in the same direction.
Numbers repeated from the previous draw. Per extra number, fourth prizes had 1.3% fewer winners than expected (95% interval: 0.8% to 1.7% fewer), and third prizes had 1.9% fewer winners (95% interval: 1.0% to 2.7% fewer). On the held-back rounds both prizes moved in the same direction.
What did not hold up
Selected on the earlier rounds but not confirmed on the later ones: numbers up to 31, an arithmetic progression of three, multiples of 7. Not selected at all: neighbouring pairs, numbers in one slip row, numbers in one slip column, touching pairs on the slip, three on a straight line of the slip.
The widely repeated idea that people lean on birthdays, and therefore on numbers up to 31, is not confirmed by this test: the count of numbers up to 31 did not pass for the third prize, and for the fourth prize it did not survive the later rounds.
A possible reading of the up-to-12 effect. A birthday can be written with a month from 1 to 12 and a day from 1 to 31. So the numbers 1 to 12 fit a birthday twice (as a month and as a day), while 13 to 31 fit it only once (as a day). If many people chose numbers from birthdays, numbers up to 12 would therefore stand out more than numbers up to 31, which is what the data show. It is only a reading: the same pattern would also arise from a general liking for small numbers, and this data cannot tell the two apart.
Show the numbers
| Feature | First part (rounds fitted) | Second part (rounds held back) | Status |
|---|---|---|---|
| numbers 31 or below | 0.6 | 0.4 | Selected, not confirmed |
| numbers 12 or below | 4.3 | 3.5 | Confirmed on the second part |
| neighbouring pairs (n, n+1) | -0.6 | -1.4 | Not selected |
| contains a 3-term arithmetic progression | 1.3 | 0.5 | Selected, not confirmed |
| numbers sharing a last digit with another | 1.4 | 1.1 | Confirmed on the second part |
| multiples of 7 | -1.5 | -0.4 | Selected, not confirmed |
| most numbers in one slip row | 0.3 | -2.3 | Not selected |
| most numbers in one slip column | 0.5 | 0.8 | Not selected |
| pairs touching on the slip (8-neighbour) | 0.4 | 0.2 | Not selected |
| three on one straight line of the slip | -0.5 | 0.4 | Not selected |
| numbers repeated from the previous draw | -1.1 | -1.5 | Confirmed on the second part |
| spread (largest minus smallest, per 10) | -2.7 | -2.6 | Confirmed on the second part |
How much does this explain? On the later rounds, using only what was learned from the earlier ones, the features account for 31% of the round-to-round variation in the fourth-prize index that the games sold and the time trend leave unexplained (18% for the third prize, 46% for the fifth). The rest is other human habits, differences between buyers that the data do not show, and plain chance.
As a robustness check, the fifth prize (millions of winners per round) points the same way for 4 of the 4 confirmed features.
What this means for a ticket holder
Putting the confirmed effects together gives each possible combination an estimated popularity relative to an average one. Across all combinations the estimate ranges from 0.90 for the least picked pattern to 1.30 for the most picked one, with a typical value of 0.99. In other words, even the most crowded patterns draw only a few tens of percent more sharers than average, and most combinations differ from each other by a few percent. The extremes are extrapolations, because winning sets that extreme are rare in the data.
Two things follow. The chance of winning is the same for every combination, so nothing here is a reason to prefer or avoid any set of numbers. And if you do win, a more crowded combination tends to share the prize with somewhat more people. The page How common is my combination? applies the confirmed effects to six numbers you enter; it runs in your browser and stores nothing.
Limits
- The data show winners, not tickets. A person can hold many tickets, and the official figures do not say which prize tickets were automatic and which were chosen by hand. Automatic tickets are random, so the effects here come from the hand-picked ones and are diluted by the rest.
- Prizes below the first match only part of the winning combination, so the results describe popular parts of combinations. Whether the same holds for complete six-number combinations cannot be checked with this data.
- The features overlap (a set with many numbers up to 12 also has a narrow spread), so each effect is the part left after the others are accounted for. The labels describe what was measured; they are not proof of why people choose as they do.
- This is an association in past data. Habits change, and a pattern that held for 1,163 rounds may not hold next year.
- Twelve features were tried. The ones that did not pass may still matter in a way this data and method cannot detect; "not confirmed" is not the same as "no effect".
Conclusion
People do not spread their tickets evenly, and the pattern is partly readable from the official numbers: the clearest signs are numbers up to 12, spread of the numbers, numbers that share a last digit, numbers repeated from the previous draw. The effects are modest, a few percent per step, and none of them changes anyone's chance of winning. What they change is the number of people a prize is shared with.
The figures on this page are refitted automatically from the latest official data (now through round 1,250). If a refit changes which features are confirmed or flips a direction, the page keeps its text until the same new result has come out three weekly refits in a row; then the page is updated automatically and the change is listed on the Corrections page. The data behind it are public: Lotto 6/45 round dataset. Related reading: the birthday-number trap, how many people win the first prize and automatic versus manual picks.
LottoFlick is independent and not an official lottery operator. Numbers are random and cannot be predicted.
Sources and verification
- LottoFlick model output used on this page (regression, validation, intervals) — computed by LottoFlick
- LottoFlick dataset: one row per round with sales, winners, features — computed by LottoFlick
- Donghaeng Lottery: Lotto 6/45 round results (public data service) — verified