Model validation report

Model probabilities are scored against the actual results of finished games. Brier score, calibration curve, recent accuracy.

Model accuracy summary

How honest are our probabilities? These are scored results against finished matches, shown next to published benchmarks.

Soccer 3-way RPS

Top tier ≤0.205 · Good ≤0.21

Pre-game Brier

Good ≤0.22 · Constant 50%: 0.25

Calibration error (ECE)

≤ 0.05

Only current-engine forecasts saved before kickoff are scored. Fewer than 30 games is provisional. Brier 0.25 is the constant-50% reference; absolute scores across market mixes are not directly comparable.

Soccer — per-league scored results

League	Matches scored	3-way RPS	Brier score	Draw calibration
EPL	13Provisional	—	0.244	no draw sample
Championship	15Provisional	—	0.223	no draw sample
LaLiga	23Provisional	—	0.199	no draw sample
Bundesliga	11Provisional	—	0.191	no draw sample
Serie A	16Provisional	—	0.204	no draw sample
Ligue 1	12Provisional	—	0.202	no draw sample
UCL	No finished matches scored yet — first results appear after 30 games.
UEL	17Provisional	—	0.234	no draw sample
MLS	26Provisional	—	0.216	no draw sample
J League	12Provisional	—	0.194	no draw sample
K League 1	8Provisional	—	0.228	no draw sample
Leagues Cup	No finished matches scored yet — first results appear after 30 games.

What RPS means: Ranked Probability Score rewards being confident when right and humble when wrong; lower is better, and strong public soccer models sit near 0.205.

Baseball, basketball, football & esports scored results

League	Games	Samples	Brier (raw)
MLB	147	12054	0.194 (0.202)Good
WNBA	27	162Provisional	0.220 (0.220)
NBA	0	0Provisional	— (—)
NFL	18	108Provisional	0.241 (0.241)
LoL	8	32Provisional	0.203 (0.209)
CS2	90	304	0.237 (0.239)Watch

Weighted by sample size across the last 20 morning grading runs. Overall calibration error (ECE) is 0.019 (raw 0.063); the benchmark is 0.05 or lower.Top tier

Per-league metric charts

RPS & Brier (lower is better)

Dashed lines are the good benchmarks (RPS 0.21, Brier 0.22); the red line at 0.25 is the base-rate baseline.

Calibration curve (3-way outcome)

No finished matches scored yet — first results appear after 30 games.

X axis is the predicted probability bucket, Y axis the observed frequency. Closer to the dashed line means more honest probabilities. Switch leagues to inspect a single league's curve.

Draw honesty

No finished matches scored yet — first results appear after 30 games.

Draws are the hardest outcome to get right. Observed league draw rates usually fall between 23% and 30%. A plain scoring distribution underrates low-scoring scorelines, so the Spolytics Engine applies its own low-score correction lifting 0-0 and 1-1. Leagues where predicted and actual diverge by more than 4pp get refit on the rolling window.

Last updated 2026-10-03 This summary is fully free. Download the raw data as CSV

Daily learning trend

Last 13 days · 12,713 graded samples · 180 refits

Every morning the previous day's finished games are graded and the per-league, per-market calibrators are refitted on those results. The lines below show how that retraining moved actual performance.

7-day rolling Brier Daily Brier (calibrated) Daily Brier (raw) Calibrators refitted that day

Brier, last 7d

0.2023

vs prior 7d -0.0009

Accuracy, last 7d

69.2%

vs prior 7d +0.3pp

Last retrain

2026-10-03

12 calibrators updated · 151 samples

ECE

11.8%

raw 9.0%

Show daily table

Date	Samples	Accuracy	Brier	Raw	ECE	Refits
2026-10-03	151	74.8%	0.2020	0.2011	11.8%	12
2026-10-02	4	75.0%	0.2135	0.2738	24.9%	0
2026-09-22	217	71.0%	0.1949	0.2025	4.3%	0
2026-09-20	1,773	68.9%	0.2033	0.2059	3.1%	0
2026-09-11	524	71.4%	0.1857	0.1990	7.7%	17
2026-09-10	1,444	67.8%	0.2102	0.2138	4.2%	15
2026-09-09	1,334	69.3%	0.2002	0.2078	4.4%	14
2026-09-08	1,038	69.5%	0.2043	0.2051	3.4%	17
2026-09-07	1,458	70.5%	0.1993	0.2020	3.5%	18
2026-09-06	1,648	67.1%	0.2099	0.2108	4.1%	19
2026-09-05	939	67.5%	0.2012	0.2051	4.2%	21
2026-09-04	799	72.6%	0.1898	0.2017	6.5%	22
2026-09-03	1,384	67.8%	0.2077	0.2084	4.5%	25

Pick weight retraining

Every night the pick-selection weights (probability, bucket calibration, sport performance) are refit on the last 30 days of graded results. The first two thirds train, the last third is held out for scoring. The probability models themselves are unchanged.

Active weights

prob 0.80 · calib 0.13 · perf 0.07

Adopted in last 45 days: 1

Last refit

2026-10-03

Improvement below threshold · previous weights kept

train 8,556 · eval 571 · top 15

Metric	Before	After	Change
Top-pick hit rate	68.4%	68.4%	0.0%p
Top-pick Brier	0.1955	0.1955	0.0000

Date	Status	Hit before	Hit after	Brier before	Brier after
2026-10-03	kept	68.4%	68.4%	0.1955	0.1955
2026-10-02	kept	68.9%	68.9%	0.1996	0.1996
2026-09-22	adopted	60.0%	64.4%	0.2154	0.1770
2026-09-21	kept	71.1%	71.1%	0.1731	0.1731
2026-09-20	kept	77.8%	77.8%	0.1612	0.1612

Scored 2026-09-26 to 2026-10-02 · 1 finished games · n = 90 markets · stored Oct 3 (Sat) · 8:00 AM UTC

Overall summary

2026-09-26 ~ 2026-10-02 · 1 finished games · 90 markets

Brier score

0.2066

0.25 = blindly guessing 50%

Accuracy

70.0%

Directional at 50% threshold

Log loss

0.599

0.693 = coin flip

ECE

10.2%

Calibration error

Equal-mass ECE (ACE) 7.0% — fixed 10-bin ECE distorts when predictions cluster.

Better than a coin flip, but not by much.

Calibration effect (before → after)

Brier 0.2047 → 0.2066
ECE 6.7% → 10.2%
Accuracy 70.0% → 70.0%
Mean predicted 36.8% → 29.7% · actual rate 36.7%

Deep validation is Pro

Summary metrics and the first calibration curve are free. Per-market curves, walk-forward validation, the daily table, and windows longer than 7 days are computed for Pro.

See Pro

$7.99/mo or $59/yr · cancel anytime

Combo pick record

Loading…

NFL featured combo record

Loading…

This report evaluates only model snapshots saved before each game. Post-game recalculations and older samples of unknown origin are excluded from training and official evaluation. It is a model evaluation, separate from the published locked-pick record, and results may be empty early in a new collection period.