◈ a fun ML prediction research project · the showcase

The final, forecast

One match left in the first 48-team World Cup. Argentina — champions, chasing back-to-back stars, Messi at 39 — against Spain, the machine that hasn't conceded a knockout half it didn't control. I built a model from public data, audited it until it stopped lying to me, and locked its number in before kickoff. This page is the fun half: the tournament, the two roads to MetLife, the players — and a forecast you can play with. The workshop half, with every wrong turn, lives on its own page.

0'Kick-off: your call, then mine

Before this page shows you a single number, it wants yours. Not to be cute — because the whole page is about what a prediction even is, and the fastest way to feel that is to make one and have to live with it. Pick a winner. Say how sure you are. The scoreboard stays dark until you do.

Locked in? Good. Now — where does a number like mine even come from? From everything that actually happened. Start where the tournament started.

15'The group stage: twelve tables, twelve stories

The first 48-team World Cup: twelve groups of four, the top two advancing plus the eight best third-placed sides. Every number below is the real thing. Argentina won Group J with a perfect nine points — a Messi hat-trick — three goals by one man in a single game — in the opener. Spain won Group H without conceding a single goal. And the tournament's sweetheart: debutants Cape Verde, who drew all three and squeezed through — the first World Cup debutant to reach the knockouts since 2010.

Twelve tables seed one bracket — and now the road narrows. From here, every match is sudden death.

30'The knockouts: one bracket, sudden death

Thirty-two survivors, five rounds, no second chances. Walk it round by round below — Argentina's road in blue, Spain's in red. Watch how differently they travelled: Spain never needed a minute past ninety; Argentina went to extra time twice — and won both. (My model, for the record, predicted every one of these matches blind before scoring itself — it called 67% of them right, the betting market 69%. The full receipts live on the workshop page.)

Two survivors. Before I ask a model about them, rewind their roads and lay them side by side — because the two stories could not be less alike.

HTHalf-time analysis: two roads to MetLife

Lay the seven-match roads side by side and the final's whole character appears. Spain cruised: never behind for long, never past ninety minutes, one goal conceded all tournament — six clean sheets (full matches without conceding a goal) for keeper Unai Simón, a tournament high. Argentina survived: twice behind against Cape Verde, 0–2 down to Egypt before the comeback, extra time twice, and a semi-final flipped in seven late minutes by two Messi crosses. Control against chaos. The cagey, careful final my model expects is exactly what happens when these two styles collide.

Campaigns are made of players. Meet the two casts — and one number about Messi that explains this whole final better than any tactics board.

60'The players: one genius vs eleven hands

Here is the player comparison that actually matters — not "who is better," but where the goals come from. Argentina's 19 goals lean overwhelmingly on one man: Messi has scored 8 and assisted 4 — a hand in roughly 63% of Argentina's entire tournament. Spain's 13 goals came from six different scorers, and they conceded once, ever. Two risk shapes: a concentrated bet against a distributed one. Contain one man and Argentina dims — but nobody has contained him yet. Spain have no single plug to pull.

Now go one level deeper — to the goals themselves. I reconstructed every goal of both campaigns from the match reports: who won the ball, who created, who finished. First the relationships — who feeds whom — and then the full explorer: all 32 goals, each scored on an honest rubric, including the question fans argue about in every bar on Earth: how much was luck — and how much only looks like luck but was manufactured?

Players produce shots; shots are where a model starts counting. Time to open the hood — the whole forecast, it turns out, rests on just two numbers.

75'How the model sees it: two numbers called λ

Strip away the drama — the comebacks, the clean sheets, the genius at 39 — and the model compresses everything you've just read into two numbers. Weighing 110 real Argentina matches and 111 real Spain matches (recent and competitive games counting most), it expects Argentina to score 1.04 goals in this final and Spain 1.17. That's it. Those two scoring rates — each called λ — are the whole engine. Because goals are counts of rare events, a century-old formula (the Poisson) unfolds each λ into the chance of scoring 0, 1, 2, 3… and crossing the two gives a price for every possible scoreline. Touch the grid: the final, as a landscape of futures.

Two λs gave me a table of every scoreline. Now stop reading the table — run it. Ten thousand times.

90'⭐ One match, ten thousand replays

This is the idea the whole page exists to hand you. When I say Spain 54%, the number is literally a count from a parallel-world machine simulation: the computer plays this exact final ten thousand times — same teams, same stadium, the scoring dice rolled fresh in each run — and I count who lifts the trophy. In roughly 5,400 of those simulated worlds Spain lift the trophy; in the other 4,600, Argentina do. First, though, answer the question below honestly — then run the machine yourself.

About three thousand of those simulated finals end the ninety minutes level — and a real final can't. Where do they go?

120'Extra time & penalties: where the draws go

A World Cup final always crowns somebody. So the 30.3% of simulated finals that end the ninety level don't vanish — they flow downhill: thirty minutes of extra time (the same two teams, at one-third rates), and if still level, the penalty shootout — one kick each from twelve yards, alternating, until someone cracks — which I price as a fair coin, on purpose. Argentina's famous shootout keeper is a real thing and a tempting one; it stays a footnote because no honest number for it exists, and this project doesn't bake anecdotes into arithmetic. Flip the switch and watch 38.2 become 53.8.

One number left to check — why should you trust the counter at all? The mirror.

FTFull time: what a forecast owes you

One last breadcrumb. Those closing prices — Spain +125, Draw +200, Argentina +255 — convert them to break-even percentages and add them up. You get 105.9%. That is not a rounding error, and what it actually is — plus the four different ways I was wrong before I was right (a model that flipped its winner on eight games, a beautiful number that turned out to be silence, the yardstick I threw away, the question nobody asked) — is the whole other half of this project.

Spain, about 54. Argentina, about 46 — and Messi. The receipts, the wrong turns, and the map of every fork: How I built it →