Model cards & calibration

These cards report the actual out-of-sample backtest metrics Golavo emits, one card per competition. They are generated from the schema-validated eval_summary*.json artifacts by scripts/build_model_cards.py — never hand-edited — so the numbers here match what CI validates. Log loss is primary. No model is a declared champion; forward evidence (the calibration record) is kept separate from these historical folds.

How to read a card

Each card lists the five deterministic candidates against the climatological baseline. Skill is 1 - model log loss / baseline log loss; its 95% interval is a seeded, fold-stratified bootstrap over held-out matches. Metrics are out-of-sample on strictly chronological folds. League strengths are not comparable across competitions — each league is modeled independently from its own pack.

Men’s senior full internationals

• Scope: Men’s senior full internationals (forward seal→score surface plus these historical test folds).
• Source snapshot: martj42/international_results ddd7249ac0c2, retrieved 2026-07-10 (CC0-1.0).
• Folds: WC2022, EURO2024, WC2026 — strictly chronological; fitting and decay selection use only rows before each fold’s cutoff.

Competition report cards (positive skill means lower log loss than climatology):

FIFA World Cup report card (2022-11-20 to 2026-07-19):

Model	Matches / folds	Log loss	Skill vs baseline (95% CI)	ECE	Fold rank
climatological (baseline)	161 / 2	1.0636	+0.0% (+0.0% to +0.0%)	0.0310	5.0 (5–5)
Elo ordinal-logit	161 / 2	0.9494	+10.7% (+5.6% to +15.4%)	0.1559	1.0 (1–1)
independent Poisson	161 / 2	1.0027	+5.7% (+0.7% to +10.6%)	0.0972	3.0 (3–3)
time-decayed Dixon-Coles	161 / 2	1.0000	+6.0% (+0.6% to +10.9%)	0.0920	2.0 (2–2)
bivariate Poisson	161 / 2	1.0027	+5.7% (+0.5% to +10.5%)	0.0972	4.0 (4–4)

Skill intervals use 2,000 seeded, fold-stratified match-bootstrap samples.

UEFA Euro report card (2024-06-14 to 2024-07-14):

Model	Matches / folds	Log loss	Skill vs baseline (95% CI)	ECE	Fold rank
climatological (baseline)	51 / 1	1.1422	+0.0% (+0.0% to +0.0%)	0.1377	5.0 (5–5)
Elo ordinal-logit	51 / 1	1.0300	+9.8% (+1.3% to +17.4%)	0.0625	4.0 (4–4)
independent Poisson	51 / 1	1.0228	+10.5% (+3.4% to +17.2%)	0.0887	2.0 (2–2)
time-decayed Dixon-Coles	51 / 1	0.9973	+12.7% (+4.5% to +19.8%)	0.0890	1.0 (1–1)
bivariate Poisson	51 / 1	1.0228	+10.5% (+3.6% to +17.7%)	0.0887	3.0 (3–3)

Skill intervals use 2,000 seeded, fold-stratified match-bootstrap samples.

Log loss by fold (primary metric; lower is better; bold = best in fold):

Model	WC2022	EURO2024	WC2026
climatological (baseline)	1.0742	1.1422	1.0565
Elo ordinal-logit	1.0157	1.0300	0.9055
independent Poisson	1.0677	1.0228	0.9598
time-decayed Dixon-Coles	1.0650	0.9973	0.9571
bivariate Poisson	1.0677	1.0228	0.9598

Every candidate beats the climatological baseline on log loss on every fold; the best model varies by fold and none is crowned a champion.

Calibration — most recent fold (WC2026):

Model	Brier	ECE	RPS
climatological (baseline)	0.6371	0.0158	0.2252
Elo ordinal-logit	0.5318	0.1603	0.1725
independent Poisson	0.5717	0.0876	0.1936
time-decayed Dixon-Coles	0.5706	0.1029	0.1925
bivariate Poisson	0.5717	0.0876	0.1936

Reliability — Elo ordinal-logit on WC2026 (Wilson 95% intervals; empty bins omitted):

Confidence bin	n	Empirical	Wilson 95%
0.3–0.4	17	0.529	[0.31, 0.74]
0.4–0.5	34	0.588	[0.42, 0.74]
0.5–0.6	28	0.786	[0.60, 0.90]
0.6–0.7	15	0.733	[0.48, 0.89]
0.7–0.8	3	0.667	[0.21, 0.94]

English Premier League

• Scope: English Premier League (historical, completed seasons only — not live).
• Source snapshot: openfootball a5dd38b3bcbe, retrieved 2026-07-11 (CC0-1.0).
• Folds: EPL2022-23, EPL2023-24, EPL2024-25 — strictly chronological; fitting and decay selection use only rows before each fold’s cutoff.

Competition report cards (positive skill means lower log loss than climatology):

English Premier League report card (2022-08-01 to 2025-06-30):

Model	Matches / folds	Log loss	Skill vs baseline (95% CI)	ECE	Fold rank
climatological (baseline)	1140 / 3	1.0623	+0.0% (+0.0% to +0.0%)	0.0298	5.0 (5–5)
Elo ordinal-logit	1140 / 3	0.9924	+6.6% (+5.0% to +8.1%)	0.0622	2.0 (1–4)
independent Poisson	1140 / 3	1.0065	+5.2% (+3.3% to +7.2%)	0.0542	1.7 (1–2)
time-decayed Dixon-Coles	1140 / 3	1.0104	+4.9% (+2.9% to +6.9%)	0.0528	3.7 (3–4)
bivariate Poisson	1140 / 3	1.0065	+5.2% (+3.3% to +7.2%)	0.0542	2.7 (2–3)

Skill intervals use 2,000 seeded, fold-stratified match-bootstrap samples.

Log loss by fold (primary metric; lower is better; bold = best in fold):

Model	EPL2022-23	EPL2023-24	EPL2024-25
climatological (baseline)	1.0495	1.0551	1.0822
Elo ordinal-logit	1.0064	0.9634	1.0075
independent Poisson	1.0250	0.9494	1.0451
time-decayed Dixon-Coles	1.0272	0.9587	1.0452
bivariate Poisson	1.0250	0.9494	1.0451

Every candidate beats the climatological baseline on log loss on every fold; the best model varies by fold and none is crowned a champion.

Calibration — most recent fold (EPL2024-25):

Model	Brier	ECE	RPS
climatological (baseline)	0.6565	0.0437	0.2358
Elo ordinal-logit	0.6025	0.0362	0.2092
independent Poisson	0.6296	0.0601	0.2229
time-decayed Dixon-Coles	0.6292	0.0536	0.2228
bivariate Poisson	0.6296	0.0601	0.2229

Reliability — Elo ordinal-logit on EPL2024-25 (Wilson 95% intervals; empty bins omitted):

Confidence bin	n	Empirical	Wilson 95%
0.3–0.4	82	0.354	[0.26, 0.46]
0.4–0.5	160	0.475	[0.40, 0.55]
0.5–0.6	85	0.588	[0.48, 0.69]
0.6–0.7	42	0.643	[0.49, 0.77]
0.7–0.8	11	0.909	[0.62, 0.98]

La Liga

• Scope: La Liga (historical, completed seasons only — not live).
• Source snapshot: openfootball a5dd38b3bcbe, retrieved 2026-07-11 (CC0-1.0).
• Folds: LALIGA2021-22, LALIGA2022-23, LALIGA2023-24 — strictly chronological; fitting and decay selection use only rows before each fold’s cutoff.

Competition report cards (positive skill means lower log loss than climatology):

La Liga report card (2021-08-01 to 2024-06-30):

Model	Matches / folds	Log loss	Skill vs baseline (95% CI)	ECE	Fold rank
climatological (baseline)	1140 / 3	1.0699	+0.0% (+0.0% to +0.0%)	0.0232	5.0 (5–5)
Elo ordinal-logit	1140 / 3	1.0032	+6.2% (+4.7% to +7.8%)	0.0555	3.3 (2–4)
independent Poisson	1140 / 3	0.9898	+7.5% (+5.5% to +9.5%)	0.0451	2.0 (1–3)
time-decayed Dixon-Coles	1140 / 3	0.9892	+7.5% (+5.6% to +9.5%)	0.0485	1.7 (1–3)
bivariate Poisson	1140 / 3	0.9898	+7.5% (+5.5% to +9.4%)	0.0451	3.0 (2–4)

Skill intervals use 2,000 seeded, fold-stratified match-bootstrap samples.

Log loss by fold (primary metric; lower is better; bold = best in fold):

Model	LALIGA2021-22	LALIGA2022-23	LALIGA2023-24
climatological (baseline)	1.0811	1.0519	1.0767
Elo ordinal-logit	1.0058	1.0006	1.0030
independent Poisson	1.0089	0.9863	0.9743
time-decayed Dixon-Coles	1.0045	0.9936	0.9694
bivariate Poisson	1.0089	0.9863	0.9743

Every candidate beats the climatological baseline on log loss on every fold; the best model varies by fold and none is crowned a champion.

Calibration — most recent fold (LALIGA2023-24):

Model	Brier	ECE	RPS
climatological (baseline)	0.6512	0.0219	0.2240
Elo ordinal-logit	0.5986	0.0711	0.1991
independent Poisson	0.5795	0.0272	0.1904
time-decayed Dixon-Coles	0.5765	0.0458	0.1894
bivariate Poisson	0.5795	0.0272	0.1904

Reliability — time-decayed Dixon-Coles on LALIGA2023-24 (Wilson 95% intervals; empty bins omitted):

Confidence bin	n	Empirical	Wilson 95%
0.3–0.4	73	0.342	[0.24, 0.46]
0.4–0.5	132	0.508	[0.42, 0.59]
0.5–0.6	84	0.512	[0.41, 0.62]
0.6–0.7	53	0.660	[0.53, 0.77]
0.7–0.8	23	0.870	[0.68, 0.95]
0.8–0.9	15	0.867	[0.62, 0.96]

Bundesliga

• Scope: Bundesliga (historical, completed seasons only — not live).
• Source snapshot: openfootball a5dd38b3bcbe, retrieved 2026-07-11 (CC0-1.0).
• Folds: BUNDESLIGA2022-23, BUNDESLIGA2023-24, BUNDESLIGA2024-25 — strictly chronological; fitting and decay selection use only rows before each fold’s cutoff.

Competition report cards (positive skill means lower log loss than climatology):

Bundesliga report card (2022-08-01 to 2025-06-30):

Model	Matches / folds	Log loss	Skill vs baseline (95% CI)	ECE	Fold rank
climatological (baseline)	918 / 3	1.0751	+0.0% (+0.0% to +0.0%)	0.0346	5.0 (5–5)
Elo ordinal-logit	918 / 3	1.0149	+5.6% (+3.9% to +7.3%)	0.0374	2.3 (1–3)
independent Poisson	918 / 3	1.0160	+5.5% (+3.4% to +7.6%)	0.0569	1.7 (1–3)
time-decayed Dixon-Coles	918 / 3	1.0198	+5.1% (+3.1% to +7.1%)	0.0588	3.3 (2–4)
bivariate Poisson	918 / 3	1.0160	+5.5% (+3.4% to +7.6%)	0.0569	2.7 (2–4)

Skill intervals use 2,000 seeded, fold-stratified match-bootstrap samples.

Log loss by fold (primary metric; lower is better; bold = best in fold):

Model	BUNDESLIGA2022-23	BUNDESLIGA2023-24	BUNDESLIGA2024-25
climatological (baseline)	1.0569	1.0754	1.0930
Elo ordinal-logit	0.9947	1.0264	1.0237
independent Poisson	0.9931	1.0197	1.0353
time-decayed Dixon-Coles	0.9970	1.0287	1.0338
bivariate Poisson	0.9931	1.0197	1.0353

Every candidate beats the climatological baseline on log loss on every fold; the best model varies by fold and none is crowned a champion.

Calibration — most recent fold (BUNDESLIGA2024-25):

Model	Brier	ECE	RPS
climatological (baseline)	0.6643	0.0662	0.2379
Elo ordinal-logit	0.6140	0.0189	0.2130
independent Poisson	0.6217	0.0886	0.2174
time-decayed Dixon-Coles	0.6206	0.0837	0.2173
bivariate Poisson	0.6217	0.0886	0.2174

Reliability — Elo ordinal-logit on BUNDESLIGA2024-25 (Wilson 95% intervals; empty bins omitted):

Confidence bin	n	Empirical	Wilson 95%
0.3–0.4	59	0.390	[0.28, 0.52]
0.4–0.5	139	0.432	[0.35, 0.51]
0.5–0.6	73	0.548	[0.43, 0.66]
0.6–0.7	30	0.733	[0.56, 0.86]
0.7–0.8	5	0.800	[0.38, 0.96]

Serie A

• Scope: Serie A (historical, completed seasons only — not live).
• Source snapshot: openfootball a5dd38b3bcbe, retrieved 2026-07-11 (CC0-1.0).
• Folds: SERIEA2021-22, SERIEA2022-23, SERIEA2023-24 — strictly chronological; fitting and decay selection use only rows before each fold’s cutoff.

Competition report cards (positive skill means lower log loss than climatology):

Serie A report card (2021-08-01 to 2024-06-30):

Model	Matches / folds	Log loss	Skill vs baseline (95% CI)	ECE	Fold rank
climatological (baseline)	1140 / 3	1.0857	+0.0% (+0.0% to +0.0%)	0.0248	5.0 (5–5)
Elo ordinal-logit	1140 / 3	1.0091	+7.1% (+5.7% to +8.6%)	0.0425	2.3 (1–4)
independent Poisson	1140 / 3	1.0071	+7.2% (+5.3% to +9.2%)	0.0427	2.7 (2–3)
time-decayed Dixon-Coles	1140 / 3	1.0040	+7.5% (+5.6% to +9.4%)	0.0453	1.3 (1–2)
bivariate Poisson	1140 / 3	1.0071	+7.2% (+5.2% to +9.2%)	0.0427	3.7 (3–4)

Skill intervals use 2,000 seeded, fold-stratified match-bootstrap samples.

Log loss by fold (primary metric; lower is better; bold = best in fold):

Model	SERIEA2021-22	SERIEA2022-23	SERIEA2023-24
climatological (baseline)	1.0900	1.0792	1.0880
Elo ordinal-logit	1.0052	1.0040	1.0182
independent Poisson	1.0065	1.0120	1.0029
time-decayed Dixon-Coles	1.0045	1.0091	0.9985
bivariate Poisson	1.0065	1.0120	1.0029

Every candidate beats the climatological baseline on log loss on every fold; the best model varies by fold and none is crowned a champion.

Calibration — most recent fold (SERIEA2023-24):

Model	Brier	ECE	RPS
climatological (baseline)	0.6587	0.0144	0.2245
Elo ordinal-logit	0.6109	0.0337	0.2010
independent Poisson	0.6003	0.0279	0.1958
time-decayed Dixon-Coles	0.5989	0.0342	0.1957
bivariate Poisson	0.6003	0.0279	0.1958

Reliability — time-decayed Dixon-Coles on SERIEA2023-24 (Wilson 95% intervals; empty bins omitted):

Confidence bin	n	Empirical	Wilson 95%
0.3–0.4	65	0.385	[0.28, 0.51]
0.4–0.5	121	0.488	[0.40, 0.58]
0.5–0.6	95	0.495	[0.40, 0.59]
0.6–0.7	68	0.662	[0.54, 0.76]
0.7–0.8	30	0.700	[0.52, 0.83]
0.8–0.9	1	0.000	[0.00, 0.79]

Ligue 1

• Scope: Ligue 1 (historical, completed seasons only — not live).
• Source snapshot: openfootball a5dd38b3bcbe, retrieved 2026-07-11 (CC0-1.0).
• Folds: LIGUE1-2022-23, LIGUE1-2023-24, LIGUE1-2024-25 — strictly chronological; fitting and decay selection use only rows before each fold’s cutoff.

Competition report cards (positive skill means lower log loss than climatology):

Ligue 1 report card (2022-08-01 to 2025-06-30):

Model	Matches / folds	Log loss	Skill vs baseline (95% CI)	ECE	Fold rank
climatological (baseline)	992 / 3	1.0734	+0.0% (+0.0% to +0.0%)	0.0298	5.0 (5–5)
Elo ordinal-logit	992 / 3	1.0198	+5.0% (+3.4% to +6.5%)	0.0485	3.0 (1–4)
independent Poisson	992 / 3	1.0123	+5.7% (+3.7% to +7.6%)	0.0543	1.3 (1–2)
time-decayed Dixon-Coles	992 / 3	1.0148	+5.5% (+3.6% to +7.4%)	0.0437	3.3 (3–4)
bivariate Poisson	992 / 3	1.0123	+5.7% (+3.9% to +7.6%)	0.0543	2.3 (2–3)

Skill intervals use 2,000 seeded, fold-stratified match-bootstrap samples.

Log loss by fold (primary metric; lower is better; bold = best in fold):

Model	LIGUE1-2022-23	LIGUE1-2023-24	LIGUE1-2024-25
climatological (baseline)	1.0747	1.0917	1.0535
Elo ordinal-logit	1.0190	1.0445	0.9960
independent Poisson	1.0216	1.0337	0.9794
time-decayed Dixon-Coles	1.0229	1.0349	0.9844
bivariate Poisson	1.0216	1.0337	0.9794

Every candidate beats the climatological baseline on log loss on every fold; the best model varies by fold and none is crowned a champion.

Calibration — most recent fold (LIGUE1-2024-25):

Model	Brier	ECE	RPS
climatological (baseline)	0.6363	0.0302	0.2358
Elo ordinal-logit	0.5938	0.0671	0.2147
independent Poisson	0.5820	0.0591	0.2099
time-decayed Dixon-Coles	0.5853	0.0626	0.2103
bivariate Poisson	0.5820	0.0591	0.2099

Reliability — independent Poisson on LIGUE1-2024-25 (Wilson 95% intervals; empty bins omitted):

Confidence bin	n	Empirical	Wilson 95%
0.3–0.4	40	0.425	[0.29, 0.58]
0.4–0.5	143	0.483	[0.40, 0.56]
0.5–0.6	76	0.645	[0.53, 0.74]
0.6–0.7	35	0.714	[0.55, 0.84]
0.7–0.8	10	0.700	[0.40, 0.89]
0.8–0.9	2	1.000	[0.34, 1.00]

Promotion criteria for challengers

A black-box challenger (e.g. gradient boosting on engineered features, including Dixon-Coles outputs) may be considered only after: (1) at least two full forward seasons of evaluation, (2) better RPS and log loss (paired bootstrap, p < 0.05), (3) no calibration regression, and (4) a feature-attribution audit. Until then it stays a lab exhibit, not a shipped model.

Full method, leakage controls, and references: Prediction methodology.