A prediction model for FIFA World Cup matches, trained on 150+ years of international football results (49,000+ matches, 1872 to today). Predictions are served from a Dixon-Coles goal model with a squad market-value prior, locked before kickoff, and scored against real results as the tournament unfolds. The harder question behind the project was whether the probabilities are sharp enough to beat a bookmaker's closing line. Tested against real closing odds from two past World Cups, the model comes out level with the line: a small edge after calibration, but well inside the margin of error, so not a demonstrated one. Seventeen other ideas that looked promising were tested and rejected out of sample, and they stay documented alongside the three that shipped. How it gets there, and where it stops, is the rest of this README.

Live demo: qnicondavid.github.io/world-cup-predictor

At a glance

Every headline claim maps to its number, where it is proven, and the honesty label it has to keep. The status words are fixed: held-out result, parity, in progress, no verdict, not adopted.

Claim	Number	Where proven	Status
Sharp enough to beat the closing line?	level: model 0.5643 vs market 0.5684 over 99 matches	Against the closing line	parity
Production model on five held-out World Cups	0.5441 multiclass Brier over 320 matches (from 0.5717)	More negative findings	held-out result
Elo engine on the same 320 matches	184/320 (57.5%), binary Brier 0.148	Track record on past World Cups	held-out result
Squad market-value prior lift	0.6123 to 0.5907 on 2022, fully out of sample	Squad market value	held-out result
Draw-transfer calibration	0.5506 to 0.5441, all five held-out	Draw-transfer calibration	held-out result
Ideas tested and rejected	17 documented, none adopted	More negative findings	not adopted
Live 2026 record (small sample)	tracked daily in the table below, still a small sample	Live results	in progress
Model vs market consensus, live 2026	the de-vigged cross-book average is ahead so far, but it is sharper than any single book, so treat the 2018/2022 parity as the calibrated read	Against the closing line	in progress

The load-bearing figure is the held-out 0.5441. The live 2026 record is still a small sample, and the closing-line row is parity, not a win. Multiclass and binary Brier sit on scales about four times apart, so the two track-record rows are not directly comparable.

Live results

A GitHub Action runs daily: it pulls fresh results, locks predictions for upcoming fixtures with the production model, scores completed ones, and rewrites the tables below automatically.

Δ is the total goal difference from the actual result (🎯 = exact), and Brier is multiclass.

Record: 71/104 picks correct (68.3%), multiclass Brier 0.503, mean goal error 2.0 (uniform guess = 0.667)

Date	Match	Winner	H/D/A %	Score (xG)	Result	Δ	Hit
Jul 19	Spain vs Argentina	Spain	46/24/29%	1-1 (1.1-1.0)	1-0	1	✅
Jul 18	France vs England	France	44/24/32%	1-1 (1.1-1.1)	4-6	8	❌
Jul 15	England vs Argentina	Argentina	33/25/42%	0-1 (1.0-1.1)	1-2	2	✅
Jul 14	France vs Spain	Spain	32/24/44%	1-1 (1.1-1.2)	0-2	2	✅
Jul 11	Norway vs England	England	22/22/57%	0-1 (0.9-1.5)	1-2	2	✅
Jul 11	Argentina vs Switzerland	Argentina	58/22/20%	1-0 (1.5-0.8)	3-1	3	✅
Jul 10	Spain vs Belgium	Spain	57/21/21%	1-0 (1.5-1.0)	2-1	2	✅
Jul 9	France vs Morocco	France	49/25/26%	1-0 (1.1-0.9)	2-0	1	✅
Jul 7	Argentina vs Egypt	Argentina	71/19/10%	1-0 (1.6-0.5)	3-2	4	✅
Jul 7	Switzerland vs Colombia	Colombia	27/23/50%	1-1 (1.0-1.3)	0-0	2	❌
Jul 6	United States vs Belgium	Belgium	30/21/49%	1-1 (1.3-1.5)	1-4	3	✅
Jul 6	Portugal vs Spain	Spain	29/23/48%	1-1 (1.1-1.3)	0-1	1	✅
Jul 5	Brazil vs Norway	Brazil	62/19/19%	1-1 (1.7-1.0)	1-2	1	❌
Jul 5	Mexico vs England	England	22/22/55%	0-1 (0.9-1.5)	2-3	4	✅
Jul 4	Canada vs Morocco	Morocco	22/34/45%	0-1 (0.7-1.1)	0-3	2	✅
Jul 4	Paraguay vs France	France	13/19/68%	0-1 (0.6-1.7)	0-1	0 🎯	✅
Jul 3	Australia vs Egypt	Australia	36/35/30%	0-0 (0.9-0.8)	1-1	2	❌
Jul 3	Argentina vs Cape Verde	Argentina	77/18/5%	2-0 (2.1-0.4)	3-2	3	✅
Jul 3	Colombia vs Ghana	Colombia	64/24/12%	1-0 (1.8-0.6)	1-0	0 🎯	✅
Jul 2	Spain vs Austria	Spain	58/25/17%	1-0 (1.7-0.8)	3-0	2	✅
Jul 2	Portugal vs Croatia	Portugal	50/27/23%	1-0 (1.5-0.9)	2-1	2	✅
Jul 2	Switzerland vs Algeria	Switzerland	40/28/32%	1-1 (1.4-1.2)	2-0	2	✅
Jul 1	England vs DR Congo	England	57/30/14%	1-0 (1.3-0.5)	2-1	2	✅
Jul 1	Belgium vs Senegal	Belgium	43/29/28%	1-1 (1.4-1.0)	3-2	3	✅
Jul 1	United States vs Bosnia and Herzegovina	United States	64/21/15%	2-0 (2.1-0.9)	2-0	0 🎯	✅
Jun 30	Ivory Coast vs Norway	Norway	30/30/40%	1-1 (1.0-1.2)	1-2	1	✅
Jun 30	France vs Sweden	France	62/22/16%	2-0 (2.0-1.0)	3-0	1	✅
Jun 30	Mexico vs Ecuador	Mexico	35/34/31%	0-0 (0.9-0.9)	2-0	2	✅
Jun 29	Brazil vs Japan	Brazil	45/29/26%	1-0 (1.4-1.0)	2-1	2	✅
Jun 29	Germany vs Paraguay	Germany	53/27/21%	1-0 (1.6-0.9)	1-1	1	❌
Jun 29	Netherlands vs Morocco	Morocco	33/32/35%	1-1 (1.0-1.0)	1-1	0 🎯	❌
Jun 28	South Africa vs Canada	Canada	19/30/50%	0-1 (0.7-1.3)	0-1	0 🎯	✅
Jun 27	Algeria vs Austria	Austria	33/30/37%	1-1 (1.3-1.3)	3-3	4	❌
Jun 27	Jordan vs Argentina	Argentina	2/12/86%	0-3 (0.5-3.7)	1-3	1	✅
Jun 27	Colombia vs Portugal	Colombia	36/30/35%	1-1 (1.3-1.3)	0-0	2	❌
Jun 27	DR Congo vs Uzbekistan	Uzbekistan	26/28/46%	1-1 (1.1-1.5)	3-1	2	❌
Jun 27	Panama vs England	England	8/19/73%	0-2 (0.7-2.5)	0-2	0 🎯	✅
Jun 27	Croatia vs Ghana	Croatia	81/15/4%	3-0 (3.1-0.5)	2-1	2	✅
Jun 26	Egypt vs Iran	Iran	22/27/51%	1-1 (1.0-1.7)	1-1	0 🎯	❌
Jun 26	New Zealand vs Belgium	Belgium	6/18/76%	0-2 (0.6-2.8)	1-5	4	✅
Jun 26	Cape Verde vs Saudi Arabia	Saudi Arabia	29/29/42%	1-1 (1.2-1.5)	0-0	2	❌
Jun 26	Uruguay vs Spain	Spain	10/21/69%	0-2 (0.7-2.3)	0-1	1	✅
Jun 26	Norway vs France	France	19/26/56%	0-1 (0.9-1.8)	1-4	4	✅
Jun 26	Senegal vs Iraq	Senegal	61/25/15%	1-0 (2.0-0.8)	5-0	4	✅
Jun 25	United States vs Turkey	Turkey	27/28/45%	1-1 (1.1-1.5)	2-3	3	✅
Jun 25	Paraguay vs Australia	Paraguay	37/30/33%	1-1 (1.3-1.3)	0-0	2	❌
Jun 25	Curaçao vs Ivory Coast	Ivory Coast	9/20/71%	0-2 (0.7-2.5)	0-2	0 🎯	✅
Jun 25	Ecuador vs Germany	Germany	33/30/38%	1-1 (1.2-1.4)	2-1	1	❌
Jun 25	Japan vs Sweden	Japan	67/22/11%	2-0 (2.3-0.7)	1-1	2	❌
Jun 25	Tunisia vs Netherlands	Netherlands	8/19/73%	0-2 (0.7-2.6)	1-3	2	✅
Jun 24	Mexico vs Czech Republic	Mexico	65/23/11%	2-0 (2.2-0.8)	3-0	1	✅
Jun 24	South Africa vs South Korea	South Korea	9/20/70%	0-2 (0.7-2.4)	1-0	3	❌
Jun 24	Canada vs Switzerland	Switzerland	34/30/36%	1-1 (1.3-1.3)	1-2	1	✅
Jun 24	Bosnia and Herzegovina vs Qatar	Bosnia and Herzegovina	52/27/21%	1-0 (1.7-1.0)	3-1	3	✅
Jun 24	Scotland vs Brazil	Brazil	11/23/66%	0-2 (0.8-2.2)	0-3	1	✅
Jun 24	Morocco vs Haiti	Morocco	78/17/5%	2-0 (2.9-0.6)	4-2	4	✅
Jun 23	Portugal vs Uzbekistan	Portugal	68/22/10%	2-0 (2.3-0.7)	5-0	3	✅
Jun 23	Colombia vs DR Congo	Colombia	76/18/6%	2-0 (2.8-0.6)	1-0	1	✅
Jun 23	England vs Ghana	England	88/10/2%	4-0 (4.0-0.4)	0-0	4	❌
Jun 23	Panama vs Croatia	Croatia	15/25/60%	0-1 (0.9-2.0)	0-1	0 🎯	✅
Jun 22	France vs Iraq	France	84/13/3%	3-0 (3.4-0.5)	3-0	0 🎯	✅
Jun 22	Norway vs Senegal	Norway	48/28/24%	1-1 (1.6-1.1)	3-2	3	✅
Jun 22	Argentina vs Austria	Argentina	71/20/9%	2-0 (2.4-0.7)	2-0	0 🎯	✅
Jun 22	Jordan vs Algeria	Algeria	17/25/58%	0-1 (0.9-1.9)	1-2	2	✅
Jun 21	Belgium vs Iran	Belgium	49/27/23%	1-1 (1.6-1.0)	0-0	2	❌
Jun 21	New Zealand vs Egypt	Egypt	22/27/52%	0-1 (1.0-1.7)	1-3	3	✅
Jun 21	Spain vs Saudi Arabia	Spain	91/8/2%	4-0 (4.5-0.4)	4-0	0 🎯	✅
Jun 21	Uruguay vs Cape Verde	Uruguay	77/18/6%	2-0 (2.8-0.6)	2-2	2	❌
Jun 20	Germany vs Ivory Coast	Germany	63/24/12%	2-0 (2.1-0.8)	2-1	1	✅
Jun 20	Ecuador vs Curaçao	Ecuador	87/11/2%	3-0 (3.9-0.4)	0-0	3	❌
Jun 20	Netherlands vs Sweden	Netherlands	68/22/10%	2-0 (2.3-0.7)	5-1	4	✅
Jun 20	Tunisia vs Japan	Japan	8/19/73%	0-2 (0.7-2.6)	0-4	2	✅
Jun 19	Scotland vs Morocco	Morocco	17/25/58%	0-1 (0.9-1.9)	0-1	0 🎯	✅
Jun 19	Brazil vs Haiti	Brazil	84/13/3%	3-0 (3.4-0.5)	3-0	0 🎯	✅
Jun 19	United States vs Australia	United States	38/30/32%	1-1 (1.4-1.2)	2-0	2	✅
Jun 19	Turkey vs Paraguay	Turkey	46/28/26%	1-1 (1.5-1.1)	0-1	1	❌
Jun 18	Czech Republic vs South Africa	Czech Republic	59/25/16%	1-0 (1.9-0.9)	1-1	1	❌
Jun 18	Mexico vs South Korea	Mexico	54/26/20%	1-0 (1.8-1.0)	1-0	0 🎯	✅
Jun 18	Switzerland vs Bosnia and Herzegovina	Switzerland	76/18/6%	2-0 (2.8-0.6)	4-1	3	✅
Jun 18	Canada vs Qatar	Canada	85/12/3%	3-0 (3.6-0.5)	6-0	3	✅
Jun 17	Portugal vs DR Congo	Portugal	76/18/6%	2-0 (2.8-0.6)	1-1	2	❌
Jun 17	Uzbekistan vs Colombia	Colombia	10/21/69%	0-2 (0.7-2.3)	1-3	2	✅
Jun 17	England vs Croatia	England	51/27/22%	1-1 (1.7-1.0)	4-2	4	✅
Jun 17	Ghana vs Panama	Panama	13/25/62%	0-2 (0.8-2.1)	1-0	3	❌
Jun 16	France vs Senegal	France	66/23/11%	2-0 (2.2-0.8)	3-1	2	✅
Jun 16	Iraq vs Norway	Norway	9/20/71%	0-2 (0.7-2.5)	1-4	3	✅
Jun 16	Argentina vs Algeria	Argentina	72/19/8%	2-0 (2.5-0.7)	3-0	1	✅
Jun 16	Austria vs Jordan	Austria	60/25/15%	1-0 (2.0-0.9)	3-1	3	✅
Jun 15	Belgium vs Egypt	Belgium	63/24/13%	2-0 (2.1-0.8)	1-1	2	❌
Jun 15	Iran vs New Zealand	Iran	65/24/11%	2-0 (2.2-0.8)	2-2	2	❌
Jun 15	Spain vs Cape Verde	Spain	93/5/2%	5-0 (5.0-0.3)	0-0	5	❌
Jun 15	Saudi Arabia vs Uruguay	Uruguay	8/20/72%	0-2 (0.7-2.5)	1-1	2	❌
Jun 14	Germany vs Curaçao	Germany	88/10/2%	4-0 (4.0-0.4)	7-1	4	✅
Jun 14	Ivory Coast vs Ecuador	Ecuador	14/25/61%	0-2 (0.8-2.0)	1-0	3	❌
Jun 14	Netherlands vs Japan	Netherlands	36/30/35%	1-1 (1.3-1.3)	2-2	2	❌
Jun 14	Sweden vs Tunisia	Sweden	42/29/29%	1-1 (1.5-1.2)	5-1	4	✅
Jun 13	Qatar vs Switzerland	Switzerland	3/12/85%	0-3 (0.5-3.6)	1-1	3	❌
Jun 13	Brazil vs Morocco	Brazil	46/28/26%	1-1 (1.5-1.1)	1-1	0 🎯	❌
Jun 13	Haiti vs Scotland	Scotland	15/25/60%	0-1 (0.8-2.0)	0-1	0 🎯	✅
Jun 13	Australia vs Turkey	Turkey	25/28/48%	1-1 (1.1-1.6)	2-0	2	❌
Jun 12	Canada vs Bosnia and Herzegovina	Canada	75/18/7%	2-0 (2.7-0.6)	1-1	2	❌
Jun 12	United States vs Paraguay	United States	36/30/34%	1-1 (1.3-1.3)	4-1	3	✅
Jun 11	Mexico vs South Africa	Mexico	75/15/9%	2-0 (2.0-0.6)	2-0	0 🎯	✅
Jun 11	South Korea vs Czech Republic	South Korea	44/23/33%	1-1 (1.2-1.2)	2-1	1	✅

Championship odds

The model's championship odds from 10,000 Monte Carlo simulations, updated 2026-10-10. They inherit the simulator's simplifications (knockout bracket paired in schedule order, games as neutral with no draws), so read them as the model's view, not a hard forecast.

#	Team	Title	Final	Semis
1	Spain	62.5%	100.0%	100.0%
2	Argentina	37.5%	100.0%	100.0%

Track record on past World Cups

Before predicting 2026, the model is validated on five World Cups it never saw during training. For each tournament it trains only on matches played before it, then predicts every match in it, the same information regime as predicting live.

The table below scores the Elo win-probability engine (its tuned configuration against the untuned baseline) with a two-outcome Brier that counts a draw as half a win. That is a coarser, different metric from the three-class multiclass Brier (around 0.54) reported for the production Dixon-Coles model elsewhere here: the two sit on scales roughly four times apart and are not directly comparable.

Tournament	Tuned model	Baseline
World Cup 2022	32/64 (50.0%), Brier 0.183	34/64 (53.1%), Brier 0.181
World Cup 2018	37/64 (57.8%), Brier 0.159	34/64 (53.1%), Brier 0.167
World Cup 2014	39/64 (60.9%), Brier 0.135	39/64 (60.9%), Brier 0.150
World Cup 2010	35/64 (54.7%), Brier 0.146	32/64 (50.0%), Brier 0.148
World Cup 2006	41/64 (64.1%), Brier 0.119	42/64 (65.6%), Brier 0.129
Combined (320)	184/320 (57.5%), Brier 0.148	181/320 (56.6%), Brier 0.155
Coin-flip reference	50%, Brier 0.250	n/a

The tuned model beats both the baseline and a coin flip across the combined 320 matches. One pattern stands out: World Cups are getting harder to predict. The binary Brier rises almost monotonically from 0.119 (2006) to 0.183 (2022); the field has genuinely tightened.

How it works

The production model is Dixon-Coles: every team gets a separate attack and defence rating, fit by weighted maximum likelihood on all of international history, with a home-advantage term, the low-score correlation correction (rho) that fixes Poisson's under-counting of 0-0 and 1-1, exponential time decay (2-year half-life), and shrinkage toward the average for rarely-seen teams. Squad market value is folded in as a prior on those ratings. From the fitted attack/defence pair the model produces a full scoreline distribution, and from that the win/draw/loss probabilities and most-likely score you see above.

data/results.csv -> MatchCsvParser -> Dixon-Coles fit (+ value prior) -> scoreline distribution -> predictions

The original engine is an Elo rating system: every team starts at 1500, and after each match ratings shift by K * (actual - expected) with K scaled by match importance and a home boost at non-neutral venues. Elo still drives the Monte Carlo title simulation and remains the baseline every goal model is measured against.

Why you can trust it

Every figure in the track record above comes from World Cups the model never trained on, predicting each match before its result, so nothing is fit to the games it is judged on. The live 2026 picks follow the same discipline.

Live predictions are written to predictions/predictions.csv before kickoff and never changed; the git history is the audit trail. A prediction made under an older model is preserved as-is and never re-locked.

Several intuitive ideas were tested honestly and did not earn their place, and they stay documented because that is why the rest can be believed:

• Rest-days advantage: no out-of-sample improvement (--rest).
• Bivariate covariance term: adds nothing over Dixon-Coles (Brier 0.573 vs 0.574).
• Temperature scaling: sharpened in-sample but made held-out 2022 worse on every metric, so none is applied.
• Annual regression toward the mean: never helped at any strength; national-team strength is more persistent than folklore suggests.

What did survive: goal-margin scaling (combined Brier 0.148 vs 0.155) and a small, tuned squad-value prior.

More negative findings from goal-model research

Every banked idea in one place first, then the write-ups. Δ Brier is the change in combined held-out multiclass Brier against the relevant baseline, where negative means lower Brier, so it would have helped; a near miss stays rejected. This table is the source the site's rejected-ideas panel reads from, so the two cannot drift apart. Full detail for each row is in the linked section below.

Idea	Δ Brier	95% CI	Verdict	Detail
Cross-confederation strength correction	-0.0050	[-0.0120, +0.0050]	rejected: worsens the expanded surface	detail
Favourite recalibration	not gated	[-0.0130, +0.0010]	not adopted	detail
Stronger recent-form nudge (lambda 0.60)	-0.0048	[-0.0096, +0.0004]	rejected: reverses on the expanded surface	detail
Opponent-adjusted recent-form residual	-0.0020	[-0.0042, +0.0001]	near miss, does not clear the gate	detail
Real published World Cup squads	+0.0026	[-0.0017, +0.0070]	not adopted	detail
Half-life retune (~3 years)	-0.0030	n/a	inside the noise floor	detail
Match-importance weighting	+0.0013	[-0.0015, +0.0041]	not adopted	detail
Bivariate covariance term	-0.0010	n/a	adds nothing	detail
Elo + Dixon-Coles ensemble blend	+0.0007	[0.0000, +0.0019]	dead	detail
Symmetric draw scaling	0.0000	n/a	no-op	detail
Lineup-weighted squad value	not gated	n/a	no signal	detail
Dynamic / state-space ratings	not gated	n/a	inside the noise floor	detail
Temperature scaling	not gated	n/a	not adopted	detail
Rest-days advantage	not gated	n/a	no out-of-sample gain	detail
Annual regression to the mean	not gated	n/a	never helped	detail
Injury / availability at kickoff	not gated	n/a	no-go (feasibility probe)	repo probe

Six of these were tested on the five held-out World Cups (320 matches, 2006-2022) and did not clear the bar:

• Elo + Dixon-Coles ensemble blend: leave-one-tournament-out cross-validation drove the learned weight to near-pure Elo in every fold. The blend scored +0.0007 worse than plain Elo on the held-out set (95% CI [0.0000, +0.0019]); the confidence interval never favors the blend. Dead.
• Symmetric draw scaling: a single multiplicative factor k on the draw probability, tuned via LOTO-CV, reverted to k=1.0 (no-op) every time, and held-out Brier did not move. Scaling every draw up or down uniformly changes nothing, because the freed probability flows back to both sides in proportion. That reads like a dead end, and for the symmetric version it is. The mistake was assuming the fix had to be symmetric: the model does over-produce draws (mean predicted 0.283 vs a 0.225 base rate), but the surplus belongs on the favourite, not split evenly. The asymmetric version of this idea, the draw-transfer under Methodology, is the one calibration change that later cleared the gate.
• Half-life retune: longer decay half-lives (~3 years) lean slightly better; pooled held-out Brier improved by roughly -0.003 (0.595 vs 0.598). That delta is inside the approximately plus-or-minus 0.015 to 0.020 noise floor implied by block-bootstrap CIs over five tournaments, so the change is not adopted as a standalone improvement.
• Match-importance weighting: down-weighting friendlies (0.5) and up-weighting World Cup finals (1.25) in the Poisson ratings fit, as fixed a-priori weights. The variant scored 0.5454 against the baseline 0.5441, a mean paired delta of -0.0013 with a 95% CI of [-0.0041, +0.0015] that spans zero, and only two of the five tournaments improved. Resolution fell rather than rose, the opposite of the intended effect. Reproduce with --importance-export, then verify.py --paired export_predictions_form.csv export_predictions_importance.csv. Not adopted.
• Stronger recent-form nudge (lambda): a leave-one-tournament-out sweep (--form-tune-loto) of the form coefficient found a clean interior optimum near lambda 0.60, well above the shipped 0.20, lowering pooled held-out Brier from 0.5441 to 0.5393 and improving four of five tournaments. But the paired block-bootstrap CI is [-0.0004, +0.0096], grazing zero, and 2022 regresses, so it just misses the gate. That 320-match near miss did not survive Phase 1. Re-gated on a pre-registered 2,180-match surface (the five World Cups plus 65 continental-final editions since 2000), the World Cup gain reverses: lambda 0.60 is 0.0064 worse than the shipped 0.20, paired 95 percent interval [+0.0026, +0.0103], entirely on the worse side, while 0.20 sits at the surface optimum and form-off ties it. The World Cup optimum was overfitting to 320 matches. Reproduce the sweep with --form-tune-loto, or the surface re-gate with --expanded-export=0.60 then verify.py --expanded-paired; full detail in research/phase1_results.md. Rejected.
• Real World Cup squads in place of the value proxy: the shipped prior sums the 26 most valuable players per nation; this used the actual squads (23 players, 26 in 2022) for all five tournaments, taken from Wikipedia and valued through the same Transfermarkt lookup. Transfermarkt's player valuations run too thin before 2014 to value those older squads fairly, so the gate ran on 2014, 2018 and 2022, where about 80 percent of each squad carries a valuation. The real squads did not beat the proxy: combined Brier 0.5605 for the proxy against 0.5631 for the real squads, a mean paired delta of -0.0026 with a 95% CI of [-0.0070, +0.0017] that spans zero, and two of the three tournaments worse. For a strong side the best players get called up anyway. In raw euros the proxy runs about a quarter to a third above the real 23, partly because a fifth of each real squad has no valuation, but the prior standardises the log of squad value, where the two correlate about 0.99, so swapping one for the other barely moves what the model sees. The unvalued players only pull the real totals down, so if anything the test is stacked toward the real squads and still finds nothing. The parsed squads are in data/wc_squads.csv (3,775 players); research/build_realsquad_values.py rebuilds the value table, and the gate swaps data/market_values_realsquad.csv in for market_values.csv before verify.py --paired. Not adopted. A related check on the value data itself, an attack-versus-defence value split and a squad-age term built from the Transfermarkt positions, was tested the same way and also found nothing that transfers: the split and age carry information the total value misses (they are not the 0.99 wall), but neither predicts the held-out outcome residual, pooled correlations near zero with signs that flip across 2014, 2018 and 2022 (research/phase2_results.md).

A seventh idea showed real structure but was absorbed by the value prior. A cross-confederation strength correction estimates a per-confederation-pair goal-difference residual from training data and applies it to inter-confederation matchups. On the Python Dixon-Coles baseline (no value prior) it improved combined Brier by roughly -0.016 (95% CI excludes zero, all five World Cups negative, gain in resolution). But re-measured on the production model through the export bridge (--verify-export, scored with verify.py --score), the gain collapsed to about -0.005 with a 95% CI of [-0.012, +0.005] that spans zero, and one tournament reversed sign. The squad-value prior and confederation strength are partly redundant, since rich squads cluster in the strong confederations, so the prior already absorbs roughly two-thirds of the effect. Phase 1 then re-gated it on a pre-registered 2,180-match surface, and there it is not merely absorbed but actively harmful: 0.0020 worse at scale 0.5 and 0.0047 worse at scale 1.0, both paired intervals excluding zero on the worse side, and worst of all on the 494 inter-confederation matches it targets, 0.0089 then 0.0209 worse. Rejected. The export bridge and verification harness (research/verify.py) built to settle this are kept, since every future idea is judged the same way; the full re-gate is in research/phase1_results.md. The in-fit version of the same idea, estimating the offsets jointly with the ratings rather than after them, was ruled out without building it: a free per-confederation-pair probability offset, which is more expressive than any in-fit rating offset, only worsens the production model out of sample (research/phase2_results.md), so there is nothing left for the in-fit fit to gain.

Three later ideas were tested the same way and also did not clear the gate. Favourite recalibration (a leave-one-tournament-out map that sharpens the favourite's probability) cut reliability but traded away resolution and reversed sign on one tournament, leaving a 95% CI of [-0.013, +0.001] that grazes zero: the temperature lesson a third time, not adopted. Dynamic / state-space ratings (a two-timescale strength blend) gave at most a noise-floor Brier change with no certifiable resolution lift. Lineup-weighted squad value (the actual starting XI valued as-of the match date, sourced from StatsBomb open data for 2018 and 2022) carried only a weak residual signal (correlation 0.06) that did not transfer between tournaments and cost resolution out of sample, so the harder 2006-2014 backfill was not pursued; the collectors live in research/fetch_lineups.py and research/lineup_value.py as a record of the attempt.

The most recent test, and the closest any rejected idea has come, is an opponent-adjusted recent-form residual (Phase 3, research/phase3_results.md). The shipped form nudge averages raw goals conceded over the last five matches, which is confounded by schedule strength; this variant measures each side's recent defence against what the fitted model expected it to concede against that opponent. On the 2,180-match expanded surface it beats the shipped raw feature at the same nudge strength by 0.0017, with a 95 percent interval of [0.0004, 0.0030] that clears zero, the first candidate to manage that, and the gain concentrates in intra-confederation matches where recent form is comparable. It still does not ship. Leave-one-tournament-out selection over the nudge grid lands on a stronger coefficient (0.40) whose interval grazes zero, and the World Cup 320 guard trips (2006 regresses past the noise floor at 0.20, 2022 blows out at higher coefficients). The opponent-adjustment is the right idea and the honest pooled improvement is real but tiny (0.5623 to 0.5617), too small to certify at the operating point the data selects. The form channel is closed with this result on the record.

Phase 3 then closed with a direct test of whether any structure remains at all. A learned residual probe (research/phase3_results.md) trained a gradient-boosted model leave-one-tournament-out on the production model's own locked predictions plus twelve pre-kickoff features: the rate gap, total goals, value gap, the opponent-adjusted form residual, confederation pairing, rest-days difference, and neutrality. Handed the production probabilities as inputs, it could not match the production model out of sample, let alone beat it: held-out multiclass Brier 0.5880 against production's 0.5623. A label-shuffle canary landed at 0.6829, near the base-rate floor and far above the probe, and per-tournament isolation held on all seventy folds, so the null is clean rather than an artifact. The attribution leaned on the rate gap, total, and value gap, the model's own core signals, and the confederation features it was handed produced no out-of-sample gain, the same verdict Phase 1 reached by a different route. There is no exploitable structure left for this model class to capture.

A final consistency check closed Phase 3. The live tracker refits the ratings before each match day, while the backtest fits once per tournament, so the headline could in principle misrepresent the live regime. Measured directly, it does not: the refit regime scores 0.5446 on the held-out World Cups against the fit-once 0.5441, a paired delta of -0.0004 with a 95 percent interval that spans zero. The two are statistically the same, because the form nudge already carries in-tournament information, so the fit-once headline is kept as the cleaner and fully reproducible number (research/phase3_results.md).

One research idea was checked and not attempted: measuring form in expected goals instead of raw goals. The only licensable xG source, StatsBomb open data, covers tournaments but carries no international friendlies or qualifiers in any year, which are exactly the matches a last-five recent-form window is built from, so the feature cannot be constructed or validated on public data. It is parked as not feasible (research/phase3_results.md).

The most ambitious data experiment was EA Sports FC player ratings as a second squad-quality signal beside market value (notes/model/EA_RATINGS_PLAN.md, research/phase2_results.md). It was the only idea in the campaign to pass the cheap kill tests: the EA overall rating correlates 0.905 with standardized log squad value and the goalkeeper rating only 0.668, so it carries information the value prior misses, and teams EA rates above their value beat the model's expectation on 2018 and 2022. Built into the prior as a default-off multi-signal blend and gated leave-one-tournament-out on a 768-match EA-covered surface (the 2018 and 2022 World Cups plus continental finals 2015 to 2023), it did not clear. The overall term worsened the surface; a separately pre-registered goalkeeper-only term was neutral to slightly worse on it. Both, though, helped the World Cups specifically, the goalkeeper term by up to 0.0038, which is the honest residue: the one signal genuinely independent of value helps World Cup prediction but does not survive dilution across the broader surface chosen for statistical power. Not shipped; the committed data aggregate and the multi-signal prior are kept for a future edition with wider coverage.

Across the campaign three changes survived this gate: the strengthened value prior, the recent-form nudge, and a draw-transfer calibration, together moving the held-out Brier from 0.5717 to 0.5441. The first two mostly bought resolution, sharper separation between outcomes. The third went back for a calibration defect the earlier symmetric tests had written off, once the decomposition made clear the leftover error was draws landing on the wrong side rather than too much draw mass overall. Everything else, the recalibration maps and the extra rating machinery, kept washing out. Only changes that add real signal or fix a real bias held up out of sample.

Methodology in depth

Draw modelling

The Elo expected score conflates winning and drawing (E = P(win) + P(draw)/2). To split it, P(draw) is estimated empirically: replaying the post-1980 internationals through the model shows the draw rate falling from ~30% between equal teams to ~2% at a 600-point rating gap. DrawModel interpolates that observed curve (rounded and lightly smoothed in the sparse high-gap bins) and splits E into explicit win/draw/loss probabilities. The --draw-curve command regenerates the curve from data and reproduces the shipped table to within about 0.015 over the ~37,400 post-1980 internationals in the current dataset. An honest limitation: the model never makes "draw" its single most likely outcome (~30% is the ceiling), so the draw model improves probabilities, not picks; bookmakers share this property.

Goal models

A goal model gives every team a separate attack and defence rating and predicts the scoreline, yielding win/draw/loss probabilities from first principles rather than from an empirical draw curve. Three live under com.david.worldcup.goals, comparable head-to-head via --goals:

• Dixon-Coles: the production model (described above).
• Bivariate Poisson: the same attack/defence fit plus a shared component that makes the two scores positively correlated.
• Elo-Poisson: the lightweight option, reusing the Elo gap and mapping it to two Poisson rates by regression.

Scored on 320 World Cup matches (2006-2022), train-before-each-tournament:

Model	Picks correct	Combined multiclass Brier
Dixon-Coles	183/320 (57.2%)	0.574
Bivariate Poisson	183/320 (57.2%)	0.573
Elo-Poisson	178/320 (55.6%)	0.575
Elo + DrawModel (baseline)	178/320 (55.6%)	0.576
Uniform reference	n/a	0.667

The goal models edge the Elo baseline modestly. The edge is uneven across tournaments, so an Elo + Dixon-Coles ensemble (averaging the two probability vectors) is also wired into --goals. This fixed average is a different construction from the learned blend banked as dead above, where leave-one-tournament-out drove the weight toward near-pure Elo; the average lands inside the noise floor of plain Dixon-Coles, so read it as a reference, not a win. The fitter itself is validated against research/goal_models.py: on data from a known Dixon-Coles process it recovers team attack strengths at correlation 0.99.

Squad market value

Elo and goal ratings are lagging: they learn strength from results. Squad market value is a leading signal. The model folds it in as a prior on the Dixon-Coles attack/defence ratings: a richer-than-average squad gets a higher attack and lower (better) defence prior, and each team's fitted rating is shrunk toward that prior. It reads data/market_values.csv (team,as_of,value_eur); lookups always take the most recent value on or before the match date, so nothing leaks from the future.

--values-tune grid-searches the weights on 2006-2018 and validates once on held-out 2022. Held-out testing through the verification harness showed the value prior was under-exploited, so a widened sweep was run; the tuned prior (globalWeight 0.40, sparseWeight 0, valueScale 0.60, roughly double the earlier setting) beats plain Dixon-Coles out of sample (multiclass Brier 0.5907 vs 0.6123 on 2022) and is now the default. Scored across all five held-out World Cups through the export bridge, it cuts the production model's combined Brier from 0.5717 to 0.5566, with the gain split between better calibration (reliability 0.0226 to 0.0136) and sharper separation (resolution 0.0914 to 0.0957). Caveats kept in view: the weights were tuned on 2006-2018, so four of the five tournaments in this 0.5717 to 0.5566 comparison overlap the tuning set and only the 2022 leg (0.6123 to 0.5907) is fully out-of-sample; four of five tournaments improve while 2010 is marginally worse; and the sparse-team lever still earns nothing. A leave-one-tournament-out re-tune (--values-tune-loto) confirms the fixed weights are not overfit: their pooled held-out Brier is 0.5566, and letting the grid re-tune per fold does slightly worse out of sample (0.5594), with no single weighting winning a majority of folds, so the single-split choice holds up. The shipped market_values.csv is the real, full Transfermarkt-derived table (933 rows, 182 teams), rebuildable byte-for-byte from the dumps in data/transfermarkt/ via research/build_market_values.py. Each snapshot is a top-26-by-value proxy squad (players of a nationality valued within two years of the date), not a published 23/26-man roster.

Recent form

Squad value and the fitted ratings move slowly; recent results do not. A FormAdjuster nudges the win/draw/loss probabilities by each side's recent defensive form, the mean goals conceded over its last 5 matches before kickoff (leakage-safe, since only prior matches count). Measured on the production model's held-out predictions through the export bridge, a conservative coefficient cuts the combined Brier from 0.5566 to 0.5506, with the gain almost entirely in resolution (sharper separation, the component the model was losing to) and improving in all five tournaments. The coefficient (0.20) was set by judgment; a later leave-one-tournament-out sweep (--form-tune-loto) found a clean interior optimum near 0.60 that lowers pooled held-out Brier to 0.5393 and improves four of five tournaments, but its paired CI [-0.0004, +0.0096] grazes zero (2022 regresses), so the stronger nudge just misses the gate on those 320 matches, and a pre-registered re-gate on a 2,180-match surface then reversed it (lambda 0.60 worse by 0.0064), so the conservative 0.20 is kept and is now the surface optimum (see negative findings). It is shipped and wired into the live tracker, so daily predictions carry it. One caveat: the nudge moves the probabilities, not the expected goals, so the most-likely-score column still comes from the raw model. That raw scoreline carries its own measured bias: leave-one-tournament-out, the model's expected total goals run about 0.45 below the actual at the 2018 and 2022 World Cups, near zero in 2006 and widening since, so the displayed expected goals and most-likely scores read low at recent tournaments. This shows up only in the scorelines on display, not in the win, draw, and loss probabilities or the Brier.

Draw-transfer calibration

The model over-produces draws. Averaged over the five held-out World Cups it puts about 0.283 on the draw where the real rate is 0.225, a bias Dixon-Coles inherits from its low-score correction. The obvious fix, scaling the draw probability down, does nothing out of sample (see the symmetric-scaling note under negative findings), so for a while this looked settled.

What the symmetric test missed showed up in the closing-line backtest. Split by match type, the model lost most to the market on moderate favourites, matches priced around 45 to 65 percent for one side. A Murphy decomposition put that loss in reliability rather than resolution and traced it to the home and draw probabilities. The draw mass was not simply too large; it was sitting between the two teams when it belonged on the favourite.

The fix moves a fixed fraction of each match's draw probability onto whichever side the model already favours, then renormalises. The fraction, about 0.21, is fit leave-one-tournament-out and stays stable across folds. Scored the same way as every other change, it cuts the held-out Brier from 0.5506 to 0.5441, improves all five tournaments (2010 only marginally), and the block-bootstrap interval on the gain clears zero. The whole improvement is reliability, the exact component the decomposition flagged. Direction is what makes it work: a band-restricted version and a symmetric split both failed the gate, and only the global, favourite-directed transfer passed it. It ships as Calibration.transferDraw and is applied in the export and the live tracker, so daily predictions carry it.

Calibration

--calibrate audits the production model on the held-out World Cups: reliability bins, log-loss, multiclass Brier and expected calibration error (ECE), plus a temperature fit. The finding: the value-stage model is mildly under-confident (ECE ≈ 0.06), but the calibration direction is not stable across tournaments, so no temperature is applied and the raw probabilities ship as-is. After the recent-form and draw-transfer steps the shipped model is better calibrated, ECE 0.026 on the same held-out matches (python research/verify.py --score research/export_predictions_form.csv). The practical consequence for any betting layer is to demand a margin of safety and size conservatively.

Against the closing line

This is the question the project was built around, so it earns a real answer rather than a mock one.

The value-betting loop is simple. --bets reads bookmaker odds, strips the overround to fair probabilities, and compares them to the model for each fixture. When an outcome's expected value clears a threshold it is flagged and sized by fractional Kelly, on a conservative default policy (a 5% edge floor, quarter-Kelly, capped at 5% of bankroll) because the calibration wobbles between tournaments.

The hard part was getting real closing odds to test against, which international football mostly lacks. A scrape of the 2018 and 2022 World Cups filled the gap. The model's held-out predictions were then scored against the de-vigged closing line on the same matches, which is the only fair comparison. Across 99 matches:

World Cups	Matches	Model Brier	Market Brier	Difference
2018, 2022	99	0.5643	0.5684	-0.0042

The model edges the line by a hair, but the margin sits inside one standard error and the block-bootstrap interval over the tournaments spans zero. The honest read is parity: after the draw-transfer calibration the model is level with the sharpest line available, leaning very slightly favourable, without a gap wide enough to call an edge. For a model built on public data, reaching the closing line is already near the ceiling.

The paper ROI made the same point from the other side. A naive run of the value policy showed a positive return, but almost all of it came from a few longshots that happened to land, and one of those was a data bug: neutral-venue matches label home and away differently across sources, and a position-wise join once paid a favourite's win at the underdog's price. Joining on the unordered team pair removed the phantom, and what remains is what the Brier numbers imply, small disagreements with the market that carry no reliable edge.

The live test runs forward. research/fetch_odds_live.py captures pre-kickoff prices, flags value bets under the same policy, and appends them to a never-edited ledger; research/settle_bets.py grades them and measures closing-line value. A verdict there needs a few hundred settled bets, so it is a long-run instrument, not a result yet. The early returns match the backtest: the first settled bets, all contrarian draw-or-underdog flags, lost, while the model's straight match picks came in.

Rest-days differential (experimental)

--rest tests whether a team with more recovery than its opponent has an edge the rating misses: it adds rating points per day of rest advantage (capped at 10 days) and measures held-out Brier against the unadjusted baseline. It needs no new data, so it is fully reproducible. As noted above, the effect did not survive out of sample.

Run it

Requires JDK 17+ and Maven.

mvn test                            # run the unit test suite
mvn compile exec:java               # replay history, print Elo top 15 + sample predictions
mvn compile exec:java -Dexec.args="--backtest"   # evaluate on the held-out World Cups
mvn compile exec:java -Dexec.args="--tune"       # hyperparameter grid search
mvn compile exec:java -Dexec.args="--track"      # lock/score predictions, update README
mvn compile exec:java -Dexec.args="--simulate"   # Monte Carlo: 10,000 tournament sims
mvn compile exec:java -Dexec.args="--upcoming"   # every fixture with win/draw/loss probs
mvn compile exec:java -Dexec.args="--predict=France,Argentina"   # any matchup
mvn compile exec:java -Dexec.args="--goals"      # goal models vs Elo: held-out comparison
mvn compile exec:java -Dexec.args="--rest"       # does a rest-days edge improve the model?
mvn compile exec:java -Dexec.args="--values"     # does squad market value improve the model?
mvn compile exec:java -Dexec.args="--values-tune" # grid-search the market-value prior weights
mvn compile exec:java -Dexec.args="--values-tune-loto" # leave-one-tournament-out re-tune of those weights
mvn compile exec:java -Dexec.args="--calibrate"  # reliability / log-loss audit + temperature fit
mvn compile exec:java -Dexec.args="--bets"       # value bets vs bookmaker odds (mock odds)
mvn compile exec:java -Dexec.args="--verify-export" # write held-out predictions to research/export_predictions_form.csv (production 0.5441) and _value.csv (value-prior stage 0.5566)
mvn compile exec:java -Dexec.args="--draw-curve"  # reproduce DrawModel's draw-rate curve from data (since 1980)

(PowerShell: quote the whole flag, e.g. mvn compile exec:java "-Dexec.args=--simulate".)

Datasets you can reuse

Three artifacts here are hard to find elsewhere and are free to reuse under the MIT license:

• docs/data/heldout.csv (320 rows) - every held-out prediction behind the headline: 320 World Cup matches (2006-2022), each with the model's win/draw/loss probabilities and the actual result, trained leave-one-tournament-out.
• data/wc2018_odds.csv + data/wc2022_odds.csv (50 + 50 rows) - de-vigged bookmaker closing lines for the 2018 and 2022 World Cups, 99 matches scored against the model. Real closing odds for international football are genuinely scarce.
• docs/data/live_market.json - the per-match 2026 model-vs-market comparison behind the live scoreboard: model and de-vigged market probabilities plus multiclass Brier for every 2026 match where a bookmaker price was captured.

Data & credits

Match data from martj42/international_results (includes scheduled 2026 fixtures, used as the prediction list). Refreshed daily by the tracker Action.

Squad market values come from the Transfermarkt community datasets (no public API exists, so download where you have network access, not from CI):

• dcaribou/transfermarkt-datasets: the best fit, a player_valuations table with dated market values plus national-team data, refreshed weekly. Aggregate player valuations to a squad total per national team per date to build market_values.csv. That is exactly what research/build_market_values.py does; the committed data/market_values.csv is its output and rebuilds byte-for-byte.
• salimt/football-datasets and the Kaggle mirror davidcariboo/player-scores are alternatives.

Player ratings for the unshipped EA experiment come from the community Kaggle dataset stefanoleone992/fifa-23-complete-player-dataset (EA Sports FC / FIFA ratings, © Electronic Arts). Only the derived per-team aggregate data/ea_ratings.csv is committed; the raw per-player dump stays out of the repo (data/ea_raw/ is gitignored) and is not among the MIT-reusable datasets listed above.