Hot numbers: we ran the chi-square test — here's what happens every time

Every lottery site on the internet has a "hot numbers" page. Ours does too — the difference is that next to the frequency chart we run the statistical test that tells you whether those frequencies mean anything. Spoiler from decades of published research: they don't. But "trust me" is not how this site works, so here is the test itself, worked by hand, followed by what happens when professionals run it on real lotteries.

What "hot" would have to mean

If a 6/49 lottery is fair, every number has the same chance of appearing in a draw: 6 of the 49 balls come out, so each number appears with probability 6/49 ≈ 0.1224 per draw. Over many draws, counts will not be perfectly equal — randomness is lumpy. The question a chi-square (χ²) test answers is: are the counts more unequal than ordinary randomness produces?

The recipe:

1. Expected count. If there have been N draws of a 6/49 game, 6N balls have been drawn, so each of the 49 numbers is expected to appear about 6N/49 times.

2. The χ² statistic. For each number, take (observed − expected)², divide by expected, and sum over all numbers:

   χ² = Σ (Oᵢ − Eᵢ)² / Eᵢ

   Big deviations in either direction inflate the statistic. Perfectly average counts give χ² = 0 (which would itself be suspicious — real randomness never lands exactly on expectation).

3. Degrees of freedom (df). With k categories whose counts must sum to a fixed total, df = k − 1. For 49 numbers, df = 48.

4. The p-value. Compare your χ² to the chi-square distribution with that df. The p-value is the probability that pure chance would produce deviations at least this large. Small p (conventionally below 0.05) says "this is more lopsided than chance usually gets"; anything else says "nothing to see here."

A fully worked toy example

Take a mini-lottery that draws one ball from 10 each night, run for 200 draws. Expected count per number: 200/10 = 20. Suppose the observed counts are:

Sum of the last column: χ² = 7.30, with df = 10 − 1 = 9. The 5% critical value for χ² with 9 degrees of freedom is 16.92; our 7.30 is nowhere near it, and the p-value is about 0.6. Translation: deviations this size (or bigger) happen roughly 60% of the time under pure chance.

Now notice what your eyes did with that table. Number 1 came up 26 times and number 9 only 14 — number 1 looks 86% "hotter" than number 9! A hot-numbers page would print those two in flames and ice cubes respectively. The χ² test looks at exactly the same data and says: this is what fair dice look like. That gap between what patterns look like and what they test as is the entire hot-numbers industry.

The subtlety with real lotteries

One honest wrinkle: in a 6/49 draw the six balls come out without replacement, so within a draw the counts are negatively correlated (if 7 came out, that's one fewer slot for 23). Pearson's statistic therefore doesn't follow a textbook χ² distribution exactly. This is not a hand-wavy excuse — it's the subject of a peer-reviewed paper, "χ² and the lottery" by Genest, Lockhart and Stephens (Journal of the Royal Statistical Society, Series D, 2002), which derives the corrected distribution and then applies it to roughly twenty years of Canada's Lotto 6/49. Their conclusion: neither Pearson's statistic nor the alternative tests they tried gave "serious grounds for suspecting a lack of uniformity." Twenty years of draws. Nothing.

The UK got the same treatment early: John Haigh's "The Statistics of the National Lottery" (Journal of the Royal Statistical Society, Series A, 1997) tested the first 96 UK draws for uniform frequencies and for the waiting times between appearances of each number, and found both entirely consistent with random draws.

This is the pattern across the literature: whenever a properly constructed test is pointed at a regulated lottery's draw history, the "hot" and "cold" numbers dissolve into exactly the lumpy-but-fair distribution the null hypothesis predicts.

We run this test live, on every game

You don't have to take 1997's word for it. JackpotMetrics runs the chi-square test continuously against the current draw history of every game we track — the result updates as draws happen:

• US Powerball randomness test
• UK Lotto randomness test
• EuroMillions randomness test
• Saturday Lotto randomness test
• Singapore TOTO randomness test
• Mark Six randomness test

Each page shows the observed counts, the expected count, the χ² statistic, the degrees of freedom and the p-value — the same numbers as the toy table above, just bigger. You can also feed any sequence you like into our randomness tester.

What would a failure even look like?

To be fair to the hot-numbers crowd, the test can fail — it's how physical bias would be caught. If a ball were dented or a machine miscalibrated, its number would drift persistently above expectation, the χ² statistic would climb draw after draw, and the p-value would collapse. Lottery operators run these tests themselves, weigh the balls, and rotate ball sets and machines precisely so that any bias is found before players could exploit it. The chi-square test isn't a formality; it's the tripwire. It just never trips on the frequencies that hot-number sellers point at, because those frequencies are ordinary sampling noise.

So: when someone shows you a frequency chart and says "23 is hot," the correct response is a single question — what's the p-value? If they don't know, they haven't tested anything. If they do know, it's almost certainly a number like 0.6, and they're hoping you won't ask.