The 2026 World Cup is over, Spain are champions, and our model's public ledger is closed. Final tally, across all 22 graded knockout matches: a 16-6 pick record (72.7% against three-way outcomes where a coin flip is 33%), a model log loss of 0.7527 against the market's 0.8111, and 16 flagged edge bets that went 8-8 for +5.6 units — a +35.9% ROI.
Every one of those numbers was published before kickoff and graded after, win or lose, on the World Cup page. That was the point: the tournament was a public tryout for a soccer model we'd only ever backtested.
It passed. So we're not shutting it down — we're giving it a bigger job. When the Premier League season opens on August 21, Predictium starts covering club soccer.
The final World Cup ledger
| Metric | Result |
|---|---|
| Matches graded | 22 |
| Pick record (modal 3-way call) | 16-6 (72.7%) |
| Advance-round picks | 16-6 |
| Model log loss | 0.7527 |
| Market log loss (20 games with a line) | 0.8111 |
| Flagged edge bets | 16 |
| Edge record | 8-8 |
| Edge units (sized to win 1u) | +5.6u |
| Edge ROI | +35.9% |
Two results from the last week are worth pausing on, because together they explain how this model makes money — and how it doesn't.
The semifinal we got wrong paid anyway. Our modal pick was France to beat England. England won. But the model had England's chances meaningfully higher than the market did, so the flagged bet was England at +275 — and it cashed a full unit. The pick column took a loss; the ledger got richer. That's not luck, it's the whole design: we bet the gap between our probability and the market's price, not our favorite box on the ticket.
The final we got right, we lost money on. We picked Spain to win — correct, 1-0 — but the value flag was Argentina at +260, and it missed. An even 8-8 edge record finishing nearly six units up is what edge betting looks like when the prices are right and the sizing is honest. If you want the bet-by-bet receipts, they're all on the ledger.
Why club soccer is the natural next step
Here's the part that makes this an easy decision rather than a pivot: the World Cup model was never really an "international tournament" model. It's Elo ratings updated match-by-match feeding a Dixon-Coles goals model — an architecture that was born in club soccer research, where the data is richest and the sample sizes are longest. We walk-forward tested it on three World Cups (192 matches, 57.3% modal accuracy, log loss ~7% better than base rates) precisely because international soccer is the hard mode: short samples, squads that meet once every four years, one-and-done knockouts.
Club soccer is the environment this architecture actually prefers:
- 38-game league seasons instead of 7-game knockout runs — more matches, faster Elo convergence, and a real distinction between being good and running hot.
- Weekly repetition. The same teams meet twice a season, in structurally similar contexts, for decades of training history.
- Deeper data. Club soccer has shot-level and expected-goals data at a richness international tournaments can't match — headroom for the same kind of iteration that took our NBA model from B1 to B13.
There are also honest new problems, and we'd rather name them now than explain them later: squad rotation across three competitions, January transfer windows that change a team mid-season, fixture congestion, and promoted sides with thin top-flight history. Those become model features over time — lineup and availability signals did more for our NBA model than any algorithm swap — but version one will be a rating-driven match model, stated plainly, graded publicly.
What we're shipping
Starting with the 2026-27 Premier League season:
- Match probabilities for every fixture — win / draw / loss, published before kickoff, updated as lineups and news land.
- Flagged edges against market prices, same convention as the World Cup: flagged only when our probability beats the price, sized to win one unit, every result graded in public.
- A season-long public ledger — pick record, log loss versus market, edge units. The World Cup page's format, running for nine months instead of one.
- Champions League coverage following shortly after the league phase begins, with more leagues as the model earns its way into them.
Same rules as everything else Predictium publishes: numbers up before the event, grades up after, misses left on the board.
We think that should be the standard for anyone publishing soccer probabilities: a model that shows its record including the losses, in units and log loss, not vibes. That's what we did for 22 World Cup matches. Doing it for 380 Premier League matches is a much bigger promise — which is exactly why it's worth making in public.
August 21. Arsenal–Coventry is the opener. Our number will be up before kickoff.
Predictions are probabilities, not guarantees — bet only what you can afford to lose, if you bet at all. Everything on Predictium is free while we build in public: predictium.ai.