People do not pick lottery numbers at random

Nobody reports which numbers they chose. No survey was taken and no operator data was used. The pattern below is recovered from draw outcomes alone — from how prizes were shared when a number came up, which is a measurement of what players had already chosen.

How often each number was chosen

against an even split of all tickets

Select any number for its estimate and how precise that estimate is.

1,389 draws since October 2015 · one game matrix

Every estimate · number, deviation against an even split, 95% interval

1+9.3%+7.1% to +11.5%

2+9.2%+7.1% to +11.4%

3+19.0%+16.8% to +21.3%

4+12.2%+9.9% to +14.5%

5+14.8%+12.5% to +17.2%

6+12.2%+10.0% to +14.4%

7+26.4%+23.9% to +29.0%

8+17.8%+15.4% to +20.2%

9+18.4%+16.0% to +20.9%

10+14.5%+12.2% to +16.9%

11+21.5%+19.1% to +23.9%

12+18.9%+16.6% to +21.2%

13+8.1%+5.7% to +10.6%

14+4.8%+2.7% to +7.0%

15+4.7%+2.6% to +6.9%

16+6.6%+4.6% to +8.7%

17+13.1%+10.8% to +15.3%

18+7.7%+5.6% to +9.8%

19+13.6%+11.3% to +15.8%

20+4.8%+2.7% to +6.9%

21+9.5%+7.5% to +11.4%

22+6.6%+4.4% to +8.8%

23+11.2%+9.2% to +13.3%

24+6.5%+4.4% to +8.6%

25+5.9%+3.7% to +8.1%

26+3.1%+0.9% to +5.4%

27+9.2%+7.2% to +11.2%

28+6.3%+4.4% to +8.3%

29+4.6%+2.5% to +6.8%

30+2.0%+0.030% to +4.1%

31+2.8%+0.8% to +4.8%

32−1.1%−2.9% to +0.8%

33+0.9%−1.0% to +2.7%

34−5.7%−7.7% to −3.7%

35−4.9%−6.8% to −2.9%

36−3.8%−5.6% to −2.1%

37−4.4%−6.2% to −2.6%

38−4.1%−6.0% to −2.1%

39−3.6%−5.5% to −1.8%

40−5.4%−7.2% to −3.5%

41−6.8%−8.7% to −4.9%

42−4.3%−6.2% to −2.3%

43−6.8%−8.6% to −4.9%

44−4.5%−6.3% to −2.7%

45−4.9%−6.7% to −3.0%

46−6.9%−8.8% to −4.8%

47−4.7%−6.5% to −2.9%

48−8.4%−10.3% to −6.5%

49−8.8%−10.8% to −6.8%

50−9.1%−10.9% to −7.3%

51−8.0%−9.9% to −6.1%

52−6.3%−8.0% to −4.4%

53−7.0%−8.7% to −5.2%

54−5.6%−7.4% to −3.6%

55−8.9%−10.8% to −7.0%

56−9.7%−11.5% to −7.9%

57−9.2%−11.0% to −7.4%

58−11.3%−13.0% to −9.4%

59−8.9%−10.6% to −7.2%

60−11.0%−12.8% to −9.2%

61−12.3%−13.8% to −10.7%

62−11.9%−13.5% to −10.2%

63−12.1%−13.8% to −10.5%

64−13.0%−14.6% to −11.4%

65−14.5%−16.2% to −12.7%

66−11.5%−13.2% to −9.7%

67−13.1%−14.8% to −11.4%

68−11.9%−13.6% to −10.1%

69−7.0%−8.7% to −5.2%

PB 1−7.0%−8.4% to −5.5%

PB 2−3.2%−4.7% to −1.6%

PB 3+4.3%+2.7% to +5.9%

PB 4+3.6%+2.1% to +5.1%

PB 5+6.4%+4.8% to +8.0%

PB 6+0.9%−0.7% to +2.6%

PB 7+12.6%+10.7% to +14.6%

PB 8+2.5%+0.7% to +4.3%

PB 9+3.4%+1.8% to +5.1%

PB 10+2.7%+1.0% to +4.5%

PB 11+6.2%+4.4% to +8.0%

PB 12+3.1%+1.4% to +4.8%

PB 13+9.6%+7.8% to +11.4%

PB 14−3.2%−4.6% to −1.7%

PB 15−2.4%−4.0% to −0.7%

PB 16−7.8%−9.5% to −6.1%

PB 17+1.4%−0.4% to +3.2%

PB 18+0.1%−1.4% to +1.6%

PB 19−2.9%−4.5% to −1.2%

PB 20−5.9%−7.4% to −4.5%

PB 21−1.9%−3.3% to −0.4%

PB 22−4.9%−6.5% to −3.2%

PB 23−2.5%−4.2% to −0.9%

PB 24−3.1%−4.6% to −1.7%

PB 25−4.8%−6.2% to −3.3%

PB 26−3.8%−5.4% to −2.3%

Low numbers were chosen more often

Across 1,389 draws, white balls 1 through 31 were chosen 16.1% more often than 32 through 49. The change at the boundary is abrupt: the three numbers below it carry 5.1 points more than the three above.

A step at 31 is real, and it is not the whole pattern. Above the boundary the decline continues: within 32–69 alone, the lower half sits 5.0 points above the upper half, and 69 is still below an even split. A calendar cannot explain the part of this that happens above 31, so "people pick their birthdays" is not what was measured.

Per-number rates are recovered from how prizes were shared when each number was drawn. The comparison runs 1-31 against 32-49 so that it does not depend on the width of the field, and it averages in log space, which is how the source analysis states it. 5 numbers are flagged as having no established direction and are shown rather than dropped.

The control

The pattern is not a property of the draws. Running the same estimator against simulated players who chose uniformly at random, on the same drawn sets, returns nothing: its largest weight anywhere is 0.0018 against 0.2345 in the real estimate, and its fit sits at the noise floor (R² 0.058). Against 5,000 shuffles the real statistic exceeded every one drawn.

The largest thing the simulated-random control produced, on the same scale as the grid above

+0.18% — against +26.4% at the strongest real number

The control re-runs the whole estimation on synthetic counts built from the real drawn sets and the real volume scale, with simulated players choosing uniformly at random. Its figures come from the source analysis for this same era.

How this was measured

Everything the figures above depend on, stated plainly. On this page the method is the finding’s credibility, so none of it is a footnote.

What the sample is

1,389 draws from October 2015 onward — the window in which one game matrix held throughout. It is not a 34-year record, and no claim here reaches back beyond October 2015. A single matrix is required because the estimator depends on the prize structure, which changes when the matrix does.

Where the numbers come from

No survey, no operator data, and nobody reporting their picks. The estimates are recovered from published draw outcomes: when a number is drawn, how the prize was shared reveals how many tickets carried it. Repeated across 1,389 draws, that recovers a rate per number.

Why the comparison stops at 49

The headline compares 1-31 against 32-49, not against the whole field of 69. Comparing against the whole field makes the figure depend on how wide the field is: the decline continues past the boundary, so a wider field drags the high group down and inflates the gap. The narrower comparison is the one that means the same thing in every era.

Uncertain and unavailable are shown, not hidden

Every estimate carries a 95% interval. Where that interval spans an even split the direction was not established, and the cell is hatched and says so in the table — 5 numbers are in that state here. Where a number has no estimate at all the cell is left empty and marked, and it is excluded from every figure on this page. No number is in that state in this sample. Neither case is filled in with a guess.

What is measured, and what is not

The pattern is an average over everyone who played across the whole window. It is not a claim about any individual, and it does not separate tickets chosen by a person from tickets the terminal chose — that share cannot be determined from outcome data, so no figure for it appears here.

What this is not

This is a measurement of what players chose. It is not advice, it does not identify numbers as better or worse to choose, and nothing here changes the chance of any outcome — every number is drawn with the same probability regardless of how many people chose it. What player choice affects is how a prize would be shared if it were won, not whether it is won — and that is not a reason to choose differently either: the figures above describe a whole population over eleven years, not any future draw, and they say nothing about what anyone else will choose tomorrow.

The dataset

Every per-number estimate and its interval, for both fields. The method note travels inside the file, so the numbers cannot be separated from how they were made.

Download the dataset

CC BY 4.0 · attribution

Someday — somedaycomes.app

Use this string verbatim wherever the licence is displayed.

This page measures what players chose. If you want the other kind of number — what a jackpot actually pays after tax where you live — see it by state.

The other half of the picture is the numbers actually drawn — the same era, measured against what chance expects.

Estimates for entertainment purposes, not tax advice. Your actual outcome depends on your full tax situation. The estimates on this page describe what players chose in past draws and say nothing about which numbers will be drawn; every number is drawn with the same probability. Someday does not sell lottery tickets and is not affiliated with Powerball, Mega Millions, or any state lottery.