Forecast generated 06:02 GMT · Sep 10, 2026 · Latest included result Sep 10, 2026
Numerate Choir Soccer Ratings

Model performance

Validation

Historical match forecasts compared with available betting odds, closing where available. Lower is better for both metrics; the market is a hard baseline — the goal is to stay close to it.

As of Sep 4, 2026

Quarterly frozen forecasts compared with available bookmaker odds (closing when available). Historical squad-value priors are excluded to avoid future information. This is not a test of the full production update cadence. See forecasts recorded before kickoff.

Season forecasts: a separate check

70 complete league-seasons. Title-position Brier: 0.0290 · Top-N: 0.0627 · Bottom-position: 0.0968. Lower is better.

17.3% of actual reconstructed finishes land in the outer predicted deciles; the reference is 20%. This diagnostic alone does not establish calibration.

Finish positions reconstructed from match points, GD and GF; bottom-position events are not official relegation outcomes. Deductions, historical qualification rules and playoff results are not modeled. Incomplete round-robin seasons are excluded.

Does wider uncertainty improve forecasts?

70 complete league-seasons, 1,378 team-seasons, checked at preseason and halfway through each schedule. These historical checks exclude current squad-value priors.

Horizon / noiseTitle BrierTop-N BrierBottom BrierOuter deciles
Preseason · 00.028960.062560.0980823.7%
Preseason · 0.120.029050.062700.0968017.3%
Midseason · 00.014890.033380.0557921.3%
Midseason · 0.120.015140.033820.0558117.5%

The 0.12 setting widens distributions. It improves preseason bottom-position forecasts in this sample, but does not show a clear title or top-N benefit; midseason distributions appear too wide by this diagnostic. League-season bootstrap intervals include zero for all comparisons except preseason bottom positions. The production setting remains 0.12 pending stronger forward evidence.

Download paired comparisons and uncertainty intervals

Forecast accuracy by season

WindowMatchesRPS modelRPS marketGapLog loss modelLog loss market
2024-25 (Aug-Oct)2,3570.2060.200+0.0061.0160.996
2024-25 (Nov-Jan)2,1950.2070.199+0.0071.0200.997
2024-25 (Feb-May)3,4430.2090.202+0.0071.0120.989
2025-26 (Aug-Oct)2,3940.2100.205+0.0051.0221.004
2025-26 (Nov-Jan)1,7070.2070.200+0.0071.0130.991
2025-26 (Feb-May)2,1990.2080.201+0.0071.0150.994
vs-538 2022-23 (Aug-Oct)2,9470.2090.204+0.0051.0140.997
vs-538 2022-23 (Nov-Feb)2,2290.2060.198+0.0081.0120.988

RPS is the ranked probability score over win/draw/loss. A positive gap means the betting market was sharper over that window.

Calibration

Forecast probabilities bucketed into bins: for each bin, how often did the predicted outcome actually happen? A well-calibrated model sits on the diagonal.

0%0%25%25%50%50%75%75%100%100%perfect calibrationPredicted probabilityObserved frequency

Hover a point for bin detail. Point size scales with the number of forecasts in the bin.

Data table
BinPred.Obs.n
3%3.9%8.6%70
8%7.9%6.0%671
13%12.8%11.0%1,639
18%17.8%15.8%3,438
23%23.0%23.3%9,603
28%27.3%27.8%15,252
33%32.3%31.0%7,056
38%37.5%37.2%5,373
43%42.4%42.4%4,601
48%47.4%48.1%3,566
53%52.3%51.9%2,571
57%57.3%59.5%1,717
63%62.3%64.4%1,125
68%67.3%69.8%724
73%72.3%75.8%463
78%77.3%79.2%336
83%82.0%86.6%157
88%87.2%81.0%42
Accuracy by domestic league
LeagueMatchesModel RPSMarket RPS
ARG19700.21400.2099
AUT13470.21020.2104
BEL18200.21010.2045
BRA16400.20630.1950
DNK13400.21240.2062
ENG19820.21190.2014
ENG214440.21890.2145
ESP19770.19990.1945
ESP211830.21350.2062
FRA18490.20400.1979
FRA28270.22180.2129
GER17950.20360.1991
GER27580.22220.2165
GRE16180.19000.1848
ITA19890.19570.1900
ITA29790.21320.2052
MEX16060.20140.1951
NED17850.19340.1887
NOR13660.20770.1977
POL15510.21530.2108
POR17680.18440.1784
SCO15900.19460.1913
SUI13620.21560.2129
SWE14030.21770.2085
TUR17930.20130.1877
USA17290.22140.2153