Lottery prediction software, tested against the only judge that matters

Lottery prediction software is a real industry with real prices — apps, desktop programs and subscription "AI predictors" charging anywhere from a few dollars to hundreds per year. The pitch varies; the mathematical situation does not. Let's set out precisely what such software claims, what a lottery draw permits, and what happens when the two are introduced to each other.

What the software claims

The genre's marketing converges on a few standard promises: the software has "analysed thousands of past draws"; it has identified "trends", "hot zones" or "winning patterns"; it predicts numbers with some spectacular-sounding accuracy — "82%" is the flavour of number that appears; and testimonials show users hitting 3-, 4- or 5-number matches. Note what is never offered: an audited, pre-registered record of predictions published before the draws they predict, scored against a chance baseline. Keep that absence in mind; it's the whole story.

What a lottery draw permits

A fair 6/49 draw selects each of C(49,6) = 13,983,816 combinations with equal probability, independent of all previous draws. Matches to any ticket follow the hypergeometric distribution — the same one on every game's odds page — and no analysis of past draws can shift it, because independent events carry zero information about each other. This isn't an assumption we're granting the lotteries out of courtesy: it is the tested finding of the peer-reviewed literature, most directly Genest, Lockhart & Stephens (2002), who examined twenty years of Canada's Lotto 6/49 — numbers singly, in pairs, in larger subsets — and found nothing distinguishable from uniform randomness, and Haigh (1997) with the same verdict on the UK National Lottery. For software to out-predict a uniform draw, the draw would have to be non-uniform. Decades of testing say it isn't. That's the entire argument, and no amount of user interface can route around it.

The baselines: what "success" looks like for free

Here is what random guessing achieves in a 6/49 game, per line — every figure derivable from the hypergeometric formula P(match k) = C(6,k) × C(43,6−k) / C(49,6):

Now run the con in reverse. Suppose a vendor has 5,000 customers each playing two "predicted" lines a week. That's 10,000 lines, of which about 10,000 × 1.86% ≈ 186 hit match-3-or-better every single week — including roughly 10 match-4s — by pure chance. A year of operation manufactures thousands of genuine, sincere, screenshot-wielding testimonials while the software adds exactly nothing. Testimonials are not evidence; they are the chance baseline wearing a customer's face.

Where "82% accuracy" numbers come from

Claims like this are rarely pure fabrication — that would be legally careless. They're measurements of the wrong thing, chosen because the right thing (jackpot odds improvement) is unmeasurable for a product that doesn't work. The recurring mechanisms:

• Predicting categories, not numbers. "We predicted the winning sum would fall between 115 and 200" — a range that, like the odd-even split, covers the majority of all combinations. Calling a base rate an accuracy is the genre's foundational move.
• Wide nets. Predict 20 numbers out of 49 and the expected overlap with the 6 winners is 20 × 6/49 ≈ 2.45 numbers. "Our pool contained 3 of the 6 winning numbers!" — at or below chance, announced as triumph.
• Backtesting on the training data. Rules fitted to past draws score brilliantly on those same draws — that's data dredging, and it evaporates on the next draw.
• Counting partial matches across many lines, per the baseline table above, and reporting hits without the denominator.

Ask one question of any accuracy claim: accuracy at predicting what, measured on draws that happened after the prediction was published? No vendor answers it.

The consumer-protection picture

Regulators mostly meet this industry at its uglier end — outright prize fraud. The US Federal Trade Commission documents fake prize and lottery scams as a standing category, with the golden rule that real lotteries never require payment to receive or "unlock" winnings; in a 2024 action the FTC banned three operators of a sweepstakes scheme that extracted fees from consumers with false promises of large cash prizes. Prediction software occupies the greyer zone — selling hope rather than stealing outright — which is why most of it survives with "entertainment purposes" fine print doing enormous legal work. We found no audited public trial of any commercial predictor beating chance; given the arithmetic above, this is not a gap in the record so much as the record itself.

Test any predictor yourself, in minutes

You don't need us — the test is trivially runnable at home:

1. Have the software issue predictions for the next ten draws — recorded now, before the draws.
2. Generate an equal number of random lines with our number generator as a control.
3. Score both against the actual results, and against the baseline table above (or simulate thousands of trials with the draw simulator to see the spread).
4. Repeat until the difference is statistically meaningful — mind the multiple-comparisons trap if you test many predictors.

The software will match the random control, which will match the hypergeometric table. Every time. And here's the final twist of the knife: even if software could find bias, using it would be self-defeating at the jackpot level — popular "predicted" numbers get played by every other customer of the same software, so any win splits among the whole user base. The only number-selection edge that survives scrutiny runs the other way entirely: picking unpopular combinations to reduce sharing, which requires no prediction at all. If you want to spend money improving your lottery outcomes, the going rate for the optimal strategy — random numbers, unpopular ranges, small stakes — is zero.