Data
Are hot and cold numbers real? Testing the full drawing record
Every lottery site publishes a hot-numbers table. Two tests compare the full drawing record with a uniform model. Comparing windows is exploratory and is not an out-of-sample forecast.
"Hot" numbers are the ones drawn most often recently; "cold" numbers are the laggards. Both tables are easy to compute and impossible to resist. The question is whether a ranking carries any information about the next drawing. The two tests below describe the historical counts. Measuring prediction requires a separate out-of-sample test, which is not reported here.
The current tables
| Game | Most drawn | Least drawn | Longest current dry spell |
|---|---|---|---|
| Mega Millions | 10 (87), 17 (84), 31 (80), 3 (79), 42 (79) | 67 (53), 65 (53), 51 (53), 45 (53), 23 (55) | ball 11, 58 drawings |
| Powerball | 21 (127), 61 (126), 64 (125), 28 (120), 23 (119) | 13 (77), 46 (83), 26 (83), 49 (84), 34 (84) | ball 51, 58 drawings |
In Powerball the gap between the top and bottom of that table is 50 appearances — ball 21 has been drawn 127 times and ball 13 only 77. Presented as a bar chart it looks like a clear signal.
Test 1: does the whole distribution deviate from fair?
The chi-square goodness-of-fit test compares all 70 (or 69) observed counts against the counts a fair machine would produce, and reports the probability of seeing a deviation at least this large by chance.
| Game | Expected per ball | Chi-square | df | p-value | Verdict |
|---|---|---|---|---|---|
| Mega Millions | 66.6 | 61.8 | 69 | 0.72 | no evidence of a departure under this test |
| Powerball | 102.7 | 79.8 | 68 | 0.16 | no evidence of a departure under this test |
A p-value of 0.72 for Mega Millions means that if every white ball were equally likely, counts at least this uneven would show up about 72% of the time. Powerball's p-value is 0.16. Under this test there is no evidence of a departure from that uniform model. A non-significant result is not proof that the game is fair.
Test 2: are the extremes more extreme than chance allows?
A fair distribution still produces a most-drawn and a least-drawn ball — someone has to come first. So the right question is not "is there a hottest number" but "is the hottest number hotter than a fair game would produce". We generated 4,000 synthetic histories per game with a uniform random generator and recorded the extremes.
| Statistic | Real record | Fair simulation | Simulated 95th percentile |
|---|---|---|---|
| Mega Millions: highest ball count | 87 | 86.1 | 93 |
| Mega Millions: hottest-minus-coldest | 34 | 37.5 | 46 |
| Powerball: highest ball count | 127 | 126.7 | 135 |
| Powerball: hottest-minus-coldest | 50 | 46.4 | 57 |
The real extremes sit right on the simulated averages. Randomness is lumpy: spreading a few thousand appearances over seventy slots reliably produces a leader several appearances clear of the field, and a straggler equally far behind. The existence of a hot number is a mathematical certainty. Its identity is noise.
Exploratory window comparison
The chi-square test and the extreme-value comparison above stay inside one historical sample. Shortening the window on either generator page — from the full matrix era to the last 400 or last 120 drawings — changes which balls sit at the top of the hot list. That change does not, by itself, show a lack of predictive power. Predictive performance needs an out-of-sample test: build the ranking on earlier drawings, then score it on later drawings that were not used to make the ranking. This page does not report that test.
So why does this site show hot and cold numbers?
Because the statistics are genuinely interesting, because seeing them measured properly is the fastest cure for believing in them, and because weighting a random draw by frequency produces lines that look plausible without changing anyone's odds. Our generators let you set the frequency, dry-spell and momentum weights to zero, which gives exact uniform randomness, and we verify that with a chi-square test in our own test suite.
If a site charges you for a hot-numbers system, the two tests above — goodness of fit, and extremes against a uniform simulation — are a starting comparison. They do not measure whether a ranking predicts later drawings; that needs an out-of-sample test. The law of independent trials is the model those comparisons use. No evidence of a departure under a test is not proof that the game is fair.