World Cup 2026 / Quantitative Research Note / Rudransh Singh

World Cup Round of 32 Prediction Model

What the data says about who wins, and how confident to be.

A team-strength model rated, backtested against real results, simulated forward 50,000 times, checked against the market, and stress-tested against an obvious economic story that turns out to be mostly wrong.

matches, 1872–2026
3-way backtest accuracy
vs naive baseline
32teams simulated live
01
Method

Five steps, each one earning the right to the next

Every stage here exists to answer one question: can the previous stage be trusted before building on top of it.

Give every team a strength score
Every team gets a number that goes up after a win and down after a loss, similar to a chess rating (this is called Elo). We then moved to a more careful version called Glicko-2, which also tracks how sure or unsure we are about that number, so a team that has barely played anyone strong doesn't get treated as confidently as a team that has been tested again and again.
Turn the score into a real percentage
A strength score on its own doesn't tell you the chance of a win, draw, or loss. We used something called a Davidson model (a small, well-known formula that takes two strength scores and splits them into three real percentages) tuned using this dataset's own actual draw rate, not just guessed.
Check it against matches that already happened
Before trusting the model on future games, we tested it on all 54 group-stage matches that had already been played, scoring win, draw, and loss as three separate possible outcomes rather than quietly dropping draws to make the score look better.
Simulate the rest of the tournament, thousands of times
We ran the whole knockout bracket 50,000 times in a row, with each run using the percentages from step 2 to decide who wins each match. Counting how often each team comes out on top across all those runs gives a real probability of winning it all, not just one guess.
Compare it to real money
Finally, we checked our numbers against real sportsbook odds and prediction-market prices. The two agree on which teams are favorites, with a real gap on exactly how big the favorites' chances are.
01b
Mistakes found and fixed

Two real bugs, caught before they could mislead anyone

Building this honestly meant checking the model's own work, twice, and reporting what was actually wrong rather than only showing the parts that worked.

Caught mid-build

Validating the rating engine line by line against the published academic worked example surfaced a real bug. Home advantage points were being permanently added to a team's rating instead of only shaping that match's win expectation, quietly inflating every rating into the thousands across 150 years of matches. Fixed and re-verified before any result downstream was trusted.

Caught on review

A second issue: rating deviation (RD), the entire reason for moving from Elo to Glicko-2, was computed and displayed but never actually used in the win/draw/loss output. Fixed so RD genuinely pulls a prediction toward a more humble split when a matchup is poorly tested.

02
Live simulation

50,000 tournaments, one bracket

Each bar is a single team's probability mass thinning out, round by round, from Round of 32 through to lifting the trophy.

Round of 16 Quarterfinal Semifinal Final Champion
02b
Try it yourself

Pick any two teams

Pull two of the 48 World Cup squads and the model breaks down the win, draw, and loss probability between them, using each team's actual rating gap and recent results as evidence, not a generic template. As Round of 32 and later results come in, this same reasoning re-grounds itself in the latest result for any team that's already played its next match, and flags any team that's been eliminated.

vs
03
Honesty check

When the model says 65%, does it mean 65%?

Every group-stage prediction bucketed by stated confidence, plotted against what actually happened. The diagonal is perfect calibration, and a model is in trouble if its dots drift far from it.

Sample sizes are small per bucket (8 to 18 matches), so read this as a direction of travel check, not a verdict.

04
Reality check

Our model against real money

Sportsbook odds and prediction-market pricing, both dated to the close of the group stage. The ranking agrees almost everywhere; the size of the numbers doesn't always.

Our model Sportsbook (implied) Prediction market
Why the gap

Part of this is a measurement artifact, not a modelling miss. The public sportsbook board only lists its handful of most traded favorites, and those alone already summed to 86% of total probability, meaning the other roughly 26 teams are implicitly compressed into the leftover 14%. That said, France specifically is rated about half of what two independent real-money sources suggest, and the gap is real rather than just a listing artifact. Tracing France's actual results shows a Nations League loss to Spain, a draw with Iceland in World Cup qualifying, and a friendly loss to Ivory Coast. The friendly is discounted in this model (0.8x weight versus full weight for competitive matches), but as the bigger surprise of the three results, it still cost France more rating than either competitive loss. The model is rating France on results alone; the market may be pricing in squad talent and tactical judgment that a results-only system has no way to see. Rank agreement is the more reliable signal here: both sources name the same six favorites, in nearly the same order.

Spain (NL): -12.9 rating Iceland (qualifying): -11.2 rating Ivory Coast (friendly): -16.2 rating
05
The development-economics question

Does national wealth buy World Cup success?

GDP per capita (PPP, World Bank) plotted against each team's current rating. A weak relationship exists, until you separate it by confederation.

UEFA: Europe CONMEBOL: South America CAF: Africa AFC: Asia CONCACAF: N./C. America, Caribbean OFC: Oceania
Finding

Look at the chart and the eye is drawn to Qatar, sitting alone in the far top right: the richest country in the entire field by a wide margin, and yet one of the worst-rated teams here. The overall correlation is real but thin (r=0.29, about 8% of variance explained), and it nearly vanishes once confederation is held constant. CONMEBOL is the clearest evidence against the "money buys success" story, with the highest average rating of any confederation sitting on by far the lowest average GDP per capita. What looked like a wealth effect is mostly a UEFA effect: Europe is both rich and historically strong at football, and the two get conflated unless you separate them.