AI and the lottery: what the studies really show

Apps, videos and websites promise that artificial intelligence can find the next winning numbers. We read the published studies, then trained two machine-learning models ourselves on decades of real draws. The result: the AI does exactly as well as a random pick, no better and no worse.

What the "studies" on AI and the lottery contain

SmileyNet (2024). In July 2024, Andreas Birk posted "SmileyNet – Towards the Prediction of the Lottery by Reading Tea Leaves with AI" on arXiv. The title says it all: a neural network is first trained on smileys to put it "in a good mood", then reads tea leaves to predict coin flips. It reports 72% accuracy, against 49% for ResNet-34 and 53% for YOLOv5, and concludes that several SmileyNets combined could win the lottery. Not a single lottery draw is tested.

Deep learning and time series (2023). An article by Samy S. Abu-Naser in the International Journal of Engineering and Information Systems (IJEAIS) applies deep learning and time-series analysis to a lottery dataset from Kaggle. Its abstract mentions the "promise and limitations" of AI, but gives no accuracy figure and no comparison with randomly chosen numbers.

A probability model (2024). On arXiv, Thando Nkomozake proposes a statistical model (Compound-Dirichlet-Multinomial) to "predict winning numbers" for 6-number, 5-number, pick-4 and pick-3 games, with a strategy said to be profitable for pick-3. The abstract describes no validation on draws that were not used to build the model, and no hit rate.

What these papers have in common: none shows the one thing that matters, a model that finds more correct numbers than a random line on draws it has never seen.

What data really reveals: players' choices

The serious lottery studies are about players, not balls. In 2016, four Dutch researchers (Wang, Potter van Loon, van den Assem and van Dolder) analysed 5.1 million combinations played online in the Dutch Lotto by 131,407 players. People favour small numbers (11 is the most popular), birthdays, numbers in the middle of the slip, sequences and patterns. Because prizes are shared between winners, a popular combination pays each winner less. That is the only thing data analysis really changes: not your chance of winning, but how much you share.

Our test: two AI models against chance

Rather than rely on other people's papers, we ran the test ourselves with our open data. Before each draw, the model only sees the past: how often each number came up in the last 10, 50 and 200 draws, how many draws since it last appeared, and whether it was drawn last time. Two standard machine-learning models (logistic regression and gradient boosting, i.e. combined decision trees) learn on the first 70% of draws and are then tested on the remaining 30%, which they have never seen. In each draw, the model plays the numbers it rates most likely. We count the correct numbers, on average per draw:

The "Chance" column is what a random line gets on average (6 × 6 / 49 in Lotto 6aus49, 5 × 5 / 50 in the other two). Both models stay within the margin of error, if anything slightly below. The same goes for "hot" and "overdue" numbers: across 8 lotteries, neither method beats chance (our data on GitHub and Kaggle lets you redo the calculation).

Why AI cannot do better

There is nothing to learn. AI finds regularities in data. But balls have no memory: each draw is independent of the previous ones, and machines and balls are tested to keep it that way. Past draws therefore contain no information about the next one. No neural network, however large, can extract information that does not exist.

It memorises the past. With a few thousand draws and millions of parameters, a network can "recall" the training draws perfectly. An accuracy measured on those draws proves nothing. Only a test on unseen draws, in time order, counts.

Always compare with chance. "The AI got 3 numbers right" sounds impressive, but in Lotto 6aus49 a random line gets 3 or more right about once in 54 draws, and in EuroMillions about once in 209. Without that comparison, a result means nothing.

The bottom line

No published study shows that AI can predict lottery numbers, and our own test confirms it performs exactly like chance. AI is useful for checking that draws are truly random, or for understanding which numbers players prefer. For picking numbers it is no better than a random generator, but it can help you avoid the combinations everyone else plays.

Can you predict lottery numbers?Patterns in chanceOur open data