Forty-one pages. That is how much of a spiral-bound exercise book I filled with three parallel columns of numbers before I was willing to trust an algorithm over my own mathematical modelling. I teach maths in Manchester, and for six months I ran a genuine EuroMillions analysis alongside a stack of marking: my own frequency picks, numbers from an AI prediction tool, and a straight random Lucky Dip as a control, logged every week in the same notebook. This is lottery probability with the fun stripped right out of it: no promises, just what three methods actually did against each other over half a year of draws.
My Year 10s think their maths teacher's evenings involve red pen and a quiet enthusiasm for quadratics. Most weeks that is true. But there is a notebook in my desk drawer at school that only a couple of people know about, and it has nothing to do with lesson planning — it is a six-month record of EuroMillions draws, tested three separate ways, kept mostly out of stubbornness.
Right then, here is why I started. I had got fed up with lottery advice built on the idea that numbers have memory: websites promising to spot "overdue" numbers, forum threads insisting a ball is due a reappearance. None of that survives contact with how the machine actually works. It has no record of what it dropped in March, any more than a coin remembers landing on heads the last ten times you flipped it. That is the gambler's fallacy, and it is behind a good share of the strategy content clogging up search results for this stuff.
Before the three-column experiment, I had already burned a stretch of evenings on tips from a YouTube channel that matched picks to numerology — birthdates reduced to single digits, so-called angel numbers, that sort of thing. It produced nothing but an odd cluster of sevens and elevens and no better a hit rate than numbers pulled off a bus timetable. That failure is a fair bit of why I wanted a proper side-by-side test rather than one more system I would half-believe in for a few months and quietly drop.
Three Columns, Six Months of Mathematical Modelling
The set-up itself was plain enough. Every week for six months, I wrote down three sets of numbers in the same notebook: a frequency pick built from the fifty most recent draws, a pick generated by the AI tool, and a Lucky Dip from the machine as an honest control. Nothing about the method was clever. The point was simply to stop arguing with myself about which approach made more sense and let six months of results settle it instead.
My own frequency chart was built the ordinary way — tally the last fifty draws, rank the numbers by how often they had shown up, pick from the top of the list. I have laid that method out properly elsewhere, so I will not repeat the mechanics here, but it is worth saying plainly that ranking numbers by past frequency tells you nothing reliable about the next draw. Past appearances and future ones are not connected. That is the whole of the argument, really, even before six months of data came in.
None of the three columns are designed to prove which numbers come up next, because that is not how independent draws work in the first place — a point that deserves its own proper treatment rather than a paragraph tucked into this one. What the experiment could show, and did, was how the three approaches differed in the shape of the numbers they produced, not in some hidden power to predict them.
Why Does "Overdue" Feel So Convincing?
The AI prediction tool I tested claims to use a neural network trained on years of historical draws to flag patterns a person would miss. I am sceptical of that framing, and I say as much to my classes whenever someone brings up "lucky" exam seats — a fair, well-run draw simply has no pattern sitting in it waiting to be found. Whether what the tool does counts as genuine pattern detection or a cleverly dressed-up randomiser is a question worth its own article, but what it was reliably good at, in my columns at least, was avoiding the choices everyone else makes.
Most people picking their own numbers gravitate toward birthdays, which caps them at 1 through 31. The EuroMillions main pool runs to 50, and the Lucky Star pool runs to 12, so sticking to dates quietly writes off a large chunk of the available numbers before you have even started. It is the same instinct behind the "hot" and "cold" numbers myth, just dressed in calendar clothing rather than frequency-chart language, and a full account of why that idea falls apart mathematically belongs elsewhere — the short version is that a number's past behaviour has no bearing on what it does next.
Here is the thing though: none of this changes your odds of winning. Every combination, birthday-shaped or not, has exactly the same chance of being drawn. What changes is what happens if you do win — a "logical" set of dates-only numbers is far more likely to be shared with a room full of strangers who had the same logic. That is really an expected-value question dressed up as a strategy one, and it deserves a longer explanation than I can fairly give it here.
My neighbour Fenella has an ongoing bet with me over all this. She reckons there is something in the patterns the draw data throws up; I keep telling her the balls have no memory and never will. Neither of us has paid up yet, and I suspect neither of us plans to.
A reader called Dariusz Wojcik emailed partway through the six months, asking whether the same side-by-side method would hold up for Polish Multi Lotek as well as EuroMillions, given the two games use entirely different pool sizes. He has a flat, no-nonsense way of putting probability questions that I rather appreciate, and the honest answer is that the mechanics carry across even where the exact odds do not.
The Gap Showed Up in Black and White
By week twelve of the six months, one column in my notebook was consistently producing a wider spread across the number pool than my own frequency-based picks — not luckier, just less clustered. Turning back through the pages to check I had not misread the run, the paper had that stiff, faintly swollen give a spiral-bound notebook gets after a few damp Manchester evenings, and the gap between the two columns sat there in plain rows of ink, hard to argue myself out of.
Six months works out to roughly twenty-six draws for each of the three methods, which is still a fairly small sample, and I am cautious about reading too much into any one column's lead over that span. A longer run might narrow the gap back down, or might not. That caution matters more in a subject where people are quick to declare a system "proven" off a handful of lucky weeks.
What the data did not show was the AI tool winning more often than the Lucky Dip control in raw hit terms — it did not, not once across the whole six months. What changed was the shape of the picks it produced, consistently spread wider across the pool than anything I built by hand from a frequency table.
The bigger picture, the one I have mapped out properly in my notebook full of patterns, is that logical-sounding human strategies are often the worst thing at the table. We spot a diagonal shape on the play slip and assume it must mean something, when the machine dropping those balls has no idea what shape they make on a piece of paper. It is the same conclusion I reached in my six-month reality check of AI lottery tools: the value is not in prediction, it is in getting human bias out of the way of your own picks.
Playing the Numbers Without Fooling Yourself
If you are going to keep playing, the practical shift is a small one. Stop chasing numbers you think are "overdue," stop clustering picks around dates, and think instead about the probability of the whole set rather than any single number inside it. An AI tool's real usefulness is not finding winners, because nothing does that before the draw happens — it is generating a spread that is mathematically identical to any other combination but far less likely to be sitting in somebody else's ticket too.
I am keeping the notebook. There is something steadying about the ritual of it between stacks of homework, even though I have stopped trying to out-guess the machine with frequency charts of my own making. Fenella still thinks I will come round to her side of the bet eventually. I doubt it, though I will keep filling in the columns regardless.
Maths cannot touch the randomness sitting inside the draw itself, and no notebook, chart or tool changes that fact. What it can do is stop you falling for advice that pretends otherwise. So if six months of columns and ink leave one lesson worth carrying forward, it is this: spread your numbers wide across the full pool instead of clustering them round dates, keep whatever you spend small and fixed in advance, and treat every single draw as a fresh, memoryless event no matter what happened the week before.