Model Report Card
How the model is doing, and improving
Graded in public
Every prediction, scored after the whistle.
Before each kickoff the model's full probability distribution is locked in. After the result we score it with the same proper scoring rules a forecasting researcher would use. No cherry-picking, no hindsight. This is the honest record.
The model is holding steady: recent Brier 0.192 matches the all-time 0.179 across 102 matches. The trend is flat — solid, not scary.
Measured from the pre-kickoff snapshots. A moving window of 10 recent matches against the all-time average. Sweeps, injuries, and a large single upset can all push this number.
In one glance
93 picks · 93 settledWin rate
51%
47/93 picks won
Profit
+12.1%
per $100 staked
Calibration
C
how honest the odds are
Scoreboard
Us vs the market vs the supercomputer.
Per-match · vs the market consensus
32 scored side-by-side| # | Forecaster | Hit rate | Brier |
|---|---|---|---|
| ★ | Market (closing line, devig) | 78% | 0.406 |
| 2 | wc26.tinjak.com (us)(32/102 covered) | 81% | 0.435 |
Lower Brier = sharper forecast. Hit rate = how often the favourite outcome won. The market consensus is the closing line averaged across every bookmaker we track and Shin-devigged. Every settled match is scored the same way for everybody. No cherry-picking.
The Opta supercomputer doesn't publish per-match win probabilities, only tournament-level outlooks. The head-to-head with Opta is below.
Tournament · vs the Opta supercomputer
48 teamsWe're more bullish on
Argentina+7.5pt
us 18% · opta 10%
Spain+6.4pt
us 22% · opta 16%
Colombia+4.7pt
us 7% · opta 2%
Belgium+2.9pt
us 5% · opta 2%
We're more cautious on
England-5pt
us 6% · opta 11%
Germany-3.5pt
us 2% · opta 6%
Uruguay-1.7pt
us 0% · opta 2%
Croatia-1.5pt
us 1% · opta 2%
Where we agree
Franceus 13% · opta 13%
Brazilus 8% · opta 8%
Portugalus 6% · opta 7%Full table · title, advance, group winner
| Team | Title | Advance | ||||
|---|---|---|---|---|---|---|
| us | opta | Δ | us | opta | Δ | |
| Argentina | 18% | 10% | +7.5 | 99%+ | 93% | +7 |
| Spain | 22% | 16% | +6.4 | 99%+ | 96% | +4 |
| England | 6% | 11% | −5 | 99%+ | 94% | +6 |
| Colombia | 7% | 2% | +4.7 | 99%+ | 76% | +24 |
| Germany | 2% | 6% | −3.5 | 99%+ | 89% | +11 |
| Belgium | 5% | 2% | +2.9 | 99%+ | 80% | +20 |
| Uruguay | 0% | 2% | −1.7 | 0% | 68% | −68 |
| Croatia | 1% | 2% | −1.5 | 99%+ | 78% | +22 |
 United States | <1% | 1% | −1.2 | 99%+ | 63% | +37 |
 Netherlands | 4% | 5% | −1.1 | 99%+ | 88% | +12 |
Our probabilities update live as group results land. Opta's are captured from their pre-tournament forecast on 2026-06-10. Title resolves at the final, advance + group winner resolve at the end of group stage.
Live · 102 matches scored
RPS
0.161
Ranked probability score. Lower is sharper; under 0.18 is strong.
Log loss
0.846
Punishes confident wrong calls hardest. Lower is better.
Brier
0.501
Mean squared error of the probabilities. Lower is better.
Calib. error
0.129
How far stated odds drift from reality. Under 0.05 is well-calibrated.
Calibration curve
On the dashed line, a stated 70% happens 70% of the time.
Tap a point to compare what the model said with what happened. On the diagonal means the odds matched reality.
By market
Match result (1X2)
102 settled
Brier
0.501
Calib. error
0.129
Over / Under 2.5
102 settled
Brier
0.244
Calib. error
0.090
Both teams to score
102 settled
Brier
0.253
Calib. error
0.117
How often each confidence level lands
Grouped by how confident the model was. A well-calibrated model wins close to what it claims.
The line is what the model said; the green bar is what actually happened. Close together = well-calibrated at that confidence level. Small samples swing — read the n.
Betting proof
Profit at flat stakes
Units at flat 1u stakes, by pick number. Red stretches are drawdowns below the previous high.
Closing line value
Tap a dot to read that pick. Below the dashed line means we locked a longer price than the market closed at.
How it improves
Every prediction is stamped with the model version that made it, so each change has to earn its keep. Lower RPS = a sharper model.
Value picks track record
Hit rate
51%
47/93
ROI
+12.1%
Avg edge
21.2%
Beating the market (CLV)
Closing Line Value compares the price we logged against the sharp closing line. It is the earliest reliable proof of a real edge, long before win-rate can tell.
Avg CLV
-3.4%
Beat the close
21%
75 picks
The headline scores use every snapshotted match, not just the value picks, so the calibration is unbiased by which bets we chose. Read the full method on How it works.