§ Data · the transparency layer

The data, in the open — and yours to query

Every probability on this site is drawn from one locked, fully traceable dataset. This page opens the whole store: the 20 raw sources behind the forecast, the coverage map that changed the result, a dictionary for all 35 engineered features, the league-strength normalisation, the pipeline lineage — and a live SQL console that runs in your browser over all 12 tables.

Raw data sources

807.9 MB across 583 files

Pipeline scripts

raw → processed → models → the locked forecast

Engineered features

per team, grouped into five families

Queryable tables

74,722+ cells, live in-browser SQL

Source · project data inventory — sizes and file counts measured from the raw data store; features and table cells counted from the published dataset.

§ 01

The source catalog

Twenty raw inputs, from a 49,445-match results spine to a 52k-quote odds archive. Each card shows what it is, who provides it, its span, its size and the role it plays in the forecast. Two sources are shown as attempted-but-deferred.

History 2

International results

martj42/international_results (GitHub)

Every men's A-international — results, goalscorers and penalty shootouts.

Span: 49,445 matches · 1872 → 2026
Size: 6.7 MB · 4 files
Role: History spine

World Football Elo

eloratings.net

Current world Elo for 244 teams plus year-end ratings back to 1950 — the published anchor we also recompute reproducibly from results.

Span: 15,383 team-years · 1950 → 2026
Size: 433.8 KB · 4 files
Role: Team-rating prior

Tournament & squad 1

WC2026 structure & squads

fixturedownload.com + Wikipedia

The official 104-fixture / 12-group / 48-team structure with venues and dates, and the 26-man squads (number, position, dob, caps, goals, club) for all 1,245 selected players.

Span: 104 fixtures · 48 squads · 1,245 players
Size: 2.0 MB · 6 files
Role: Tournament & squad

Squad quality 3

Transfermarkt squad values

Transfermarkt

National-team squad market values for all 48 teams (2,278 players × value, age, caps, goals) — England €1.88bn down to Jordan €16m — plus the 2016→2024 value history.

Span: 48/48 teams · 2,278 players · value history 2016 → 2024
Size: 105.6 KB · 3 files
Role: Squad quality (primary)

Coaches

Wikipedia / federation pages

Each nation's coach, their nationality, the foreign-coach flag and appointment date — feeds the coaches page and every dossier header.

Span: 48/48 coaches
Size: —
Role: Bench

SoFIFA / EA FC ratings

partial

SoFIFA / EA FC (via soccerdata)

Club EA-FC ratings (96 clubs) as a per-player quality reference. Player-level ratings proved redundant — fully superseded by Transfermarkt value.

Span: 96 clubs (club reference)
Size: 41.6 KB · 2 files
Role: Club reference

Club form & events 4

Understat club xG

Understat (via soccerdata)

Club-level expected goals (xg, np_xg, xg_chain, xg_buildup) for the top-5 European leagues, extended back to 2016-17 so each backtest tournament has its club season.

Span: 8,908 matches · 27,177 player-seasons (top-5 only)
Size: 11.2 MB · 3 files
Role: Player / club form

API-Football global club stats

API-Football

Global club appearances, minutes and goals across 107 leagues in 68 countries — the genuinely worldwide club panel that de-biased the squad layer away from a Europe-only feed.

Span: 107 leagues · 68 countries · 4,471 club-seasons
Size: 1.3 MB · 6 files
Role: Player / club form (global)

StatsBomb open data

statsbomb/open-data (GitHub)

Full event streams and lineups for recent major tournaments (WC2018/2022, Euro2020/2024, Copa2024) — the public alternative to licensed Opta/Wyscout event data.

Span: 262 matches · ~530 event/lineup files
Size: 771.3 MB · 530 files
Role: Event / player

FBref season stats

FBref (via soccerdata)

Tournament schedules and player/team season stats (goals, assists, shots). Season tables carry no xG — that is derived from StatsBomb instead.

Span: schedule + season stats · 34/48 nations resolve
Size: 556.9 KB · 4 files
Role: Player performance

Validation & backtest 2

Bookmaker odds

the-odds-api (Pro)

WC2026 outright and full match markets (h2h, totals, spreads), plus the out-of-time closing-odds consensus for the 152 backtest matches — the de-vigged market benchmark.

Span: WC2026 markets + 52,028 historical quotes · 3 backtest tournaments
Size: 7.3 MB · 9 files
Role: Validation anchor

Historical squads (backtest)

Wikipedia (past tournaments)

As-of squads for five past tournaments (WC2018/2022, Euro2020/2024, Copa2024) so every engineered feature can be rebuilt leakage-safe for the out-of-time backtest.

Span: 3,226 players · 128 team-tournaments
Size: 250.7 KB · 1 file
Role: Backtest enablement

Context & environment 6

World Bank macro

World Bank API

GDP, GDP-per-capita and population for 261 countries — the macro prior and the GDP-growth context features.

Span: 261 countries · 1990 → 2024
Size: 565.2 KB · 1 file
Role: Country prior

Venues & environment

curated venues + open-elevation + open-meteo

The 16 host venues with coordinates, elevation, June–July heat index and a full travel-distance matrix — hottest Houston 47°C, highest Mexico City 2,287m.

Span: 16 venues · per-team travel / heat / altitude
Size: 6.3 KB · 3 files
Role: Environment (§7)

US Census — diaspora & language

US Census (ACS B05006 / B16001)

Foreign-born population by origin and home-language speakers in the United States — the novel quasi-home-support and host-language-familiarity signals.

Span: foreign-born 38/48 · home-language 25/48
Size: 2.7 KB · 2 files
Role: Diaspora / language

UN migrant-stock

UN DESA migration

Bilateral international migrant-stock — the broader diaspora mechanism behind quasi-home support, complementing the US-specific census layer.

Span: bilateral migrant-stock matrix
Size: 6.1 MB · 1 file
Role: Diaspora (global)

Government sport spend (COFOG)

partial

Eurostat COFOG

Government recreation-and-sport spending as a share of GDP and its trend — an investment-context control, available for EU/EEA nations only.

Span: 12/48 teams (EU/EEA only)
Size: 15.8 KB · 1 file
Role: Investment context

Player-club climate

open-meteo (per club city)

The home-climate each player is acclimatised to, aggregated to a squad acclimatisation temperature and differenced against host-venue heat.

Span: per-player club climate → squad acclimatisation
Size: 54.9 KB · 1 file
Role: Acclimatisation

Attempted · deferred 2

Injuries / availability

empty

Transfermarkt injury pages

All 48 national-team injury pages were scraped, but returned 0 parseable entries three days before kickoff. Availability is already encoded in the finalised 26-man squads (injured players omitted).

Span: 0 parseable rows — re-runnable mid-tournament
Size: 230 B · 2 files
Role: Attempted · empty

FIFA ranking history

deferred

inside.fifa.com

Deferred — the page loads its ranking history client-side, and Elo is the stronger, fully-reproducible state variable used as the primary prior.

Span: deferred — Elo substitutes
Size: —
Role: Deferred

How the catalog is built (and why two sources are deferred)

Sizes, file counts and spans are measured directly from the data store at build time, so the catalog cannot drift from the data it describes.

Two sources are shown as attempted-but-deferred: the FIFA-ranking history, because Elo is the stronger, fully-reproducible rating prior; and the injury feed, which returned no usable rows three days before kick-off — availability is instead encoded in the finalised 26-man squads (injured players were simply omitted). Both are re-runnable mid-tournament.

§ 02

Global club coverage — the extension behind the squad model

Public club-performance feeds cover the top-5 European leagues — and top-5 presence is itself a wealth proxy, so models built on them quietly learn “non-European ⇒ weak”. This project extends the squad layer to a genuinely global club panel (API-Football, 107 leagues), lifting coverage everywhere outside Europe and flipping the gradient-boosted model from worst to best-non-market.

Fig. V10 World choropleth · share of each squad with matched club data

Squad club-data coverage, before and after the global panel

Toggle between BEFORE (Understat top-5-Europe) and AFTER (global API-Football). Before, almost everything outside Western Europe is pale — not because those squads are weak, but because their players play in leagues the feed never saw. After, the map fills in. The map is interactive (drag, zoom); the same numbers are in the table below.

Map loads when it scrolls into view. The confederation bars and the per-team table below carry the same finding without it.

England and Scotland share one shape on the map (both are GBR in ISO-3166); it is tinted with England’s coverage. Cabo Verde and Curaçao are small but present. The table below lists every nation individually.

Overall coverage rises from 48% to 85% — and the lift is largest exactly where the old feed was blind: OFC 15%→92%, AFC 17%→71%. UEFA, already well-covered, barely moves.

Source · Understat (top-5 Europe) vs API-Football (global club coverage) · world boundaries: Natural Earth

Fig. V10b Before → after, weakest-covered confederation first

Coverage uplift by confederation — biggest gains off-UEFA

OFC · 1 team +77pt

15 92%

AFC · 9 teams +54pt

17 71%

CONCACAF · 6 teams +50pt

30 81%

CAF · 10 teams +44pt

45 88%

CONMEBOL · 6 teams +30pt

54 85%

UEFA · 16 teams +19pt

73 92%

All 48 48% → 85%

before (top-5 Europe) after (global panel)

The gain is not a uniform shift — it lands precisely where top-5 European feeds see least. The confederations they under-cover (OFC, AFC, CONCACAF) gain the most; UEFA, already near-complete, gains least.

Source · Understat (top-5 Europe) vs API-Football (global club coverage) — mean squad coverage per confederation

Every nation, before → after

The full per-team coverage, biggest gainer first. South Africa, Qatar and Saudi Arabia were nearly invisible to a top-5-only feed (one or zero players); the global panel sees almost their entire squad. The two UEFA teams at the bottom dip a touch because the global match window is stricter than Understat’s season — coverage, not invention.

Nation	In top-5	Before	After	Uplift
Qatar	1	4%	100%	+96pt
South Africa	1	4%	96%	+92pt
Saudi Arabia	1	4%	92%	+88pt
New Zealand	4	15%	92%	+77pt
Korea Republic	4	15%	85%	+70pt
Egypt	4	15%	85%	+70pt
Panama	2	8%	77%	+69pt
Cabo Verde	4	15%	81%	+66pt
Mexico	8	31%	92%	+61pt
Czechia	10	38%	96%	+58pt
Canada	6	24%	76%	+52pt
Türkiye	8	31%	81%	+50pt
Uzbekistan	0	0%	50%	+50pt
Morocco	12	46%	96%	+50pt
Curaçao	5	19%	65%	+46pt
IR Iran	4	15%	58%	+43pt
Haiti	9	35%	77%	+42pt
Paraguay	9	35%	77%	+42pt
Jordan	0	0%	42%	+42pt
Bosnia and Herzegovina	12	46%	85%	+39pt
Australia	11	42%	81%	+39pt
Ecuador	8	31%	69%	+38pt
Algeria	13	50%	88%	+38pt
Uruguay	13	50%	85%	+35pt
Iraq	3	12%	46%	+34pt
Brazil	16	62%	96%	+34pt
USA	17	65%	96%	+31pt
Norway	17	65%	96%	+31pt
Tunisia	9	35%	65%	+30pt
Colombia	15	58%	85%	+27pt
Côte d'Ivoire	18	69%	96%	+27pt
Japan	16	62%	88%	+26pt
Scotland	16	62%	88%	+26pt
Sweden	16	62%	85%	+23pt
Senegal	20	77%	100%	+23pt
Ghana	17	65%	88%	+23pt
Congo DR	18	69%	88%	+19pt
Croatia	21	81%	96%	+15pt
Belgium	20	77%	92%	+15pt
Austria	21	84%	96%	+12pt
Spain	23	88%	100%	+12pt
Portugal	21	81%	92%	+11pt
England	23	88%	96%	+8pt
France	22	85%	92%	+7pt
Argentina	23	92%	96%	+4pt
Switzerland	23	88%	92%	+4pt
Germany	25	96%	96%	−0pt
Netherlands	25	96%	88%	−8pt

Source · Understat (top-5 Europe) vs API-Football (global club coverage). Sort any column; the bars show the share of each squad with matched club data.

§ 03

The feature dictionary

The 35 engineered features the models read, grouped into five families — history, squad quality, form & fitness, context, and the decoupling g. Each has a plain definition, its source, its observed range and a distribution sparkline. (The gradient-boosted model builds match-level differences x_i − x_j from this panel at train time.)

History (h) 8 features

Current Elo rating

Current Elo rating (eloratings.net method, end-2026).

1,421 med 1779.50 2,155

src Elo ratings (eloratings.net method)elo_now

Elo change over 5 years

Elo change over last 5 years.

-242 med 32 194

src Elo ratings (eloratings.net method)elo_change_5y

Elo change over 10 years

Elo change over last 10 years.

-160 med 53.50 338

src Elo ratings (eloratings.net method)elo_change_10y

Linear Elo slope per year (10y)

Linear Elo slope per year over 10y.

-25 med 5.35 31.90

src Elo ratings (eloratings.net method)elo_slope_10y

Points per game, last 15 NT matches

National-team points-per-game, last 15 matches.

0.933 med 1.87 2.47

src International match results, 1872–2026nt_last15_ppg

Goal difference per game, last 15

National-team goal-difference per game, last 15.

-0.530 med 0.835 2.60

src International match results, 1872–2026nt_last15_gd_pg

Win rate, last 15 NT matches

National-team win rate, last 15.

0.200 med 0.533 0.800

src International match results, 1872–2026nt_last15_winrate

NT recent-form trend

Recent NT form trend (slope).

-1.70 med -0.350 1.30

src International match results, 1872–2026nt_form_trend

Squad quality (s) 9 features

Mean squad age

Mean squad age.

23.70 med 25.60 27.60

src Official squad announcementsmean_age

Share of squad aged 25-29 (peak)

Share of squad aged 25–29 (peak).

0.200 med 0.387 0.526

src Official squad announcementspct_peak_25_29

# players over 30

Players over 30.

1 med 6 15

src Official squad announcementsn_over30

# players under 23

Players under 23.

1 med 12 31

src Official squad announcementsn_under23

Mean caps (experience)

Mean international caps.

12 med 24.75 43.60

src Official squad announcementsmean_caps

Largest single-club bloc

Largest same-club bloc size.

1 med 3 10

src Official squad announcementslargest_club_bloc

# distinct clubs in squad

Distinct clubs in squad.

8 med 22 26

src Official squad announcementsn_distinct_clubs

Total squad market value (EUR)

Total squad market value (Transfermarkt, EUR).

16M med €0.29bn €1.88bn

src Transfermarktsquad_value_eur

Squad-value growth (8y CAGR)

Squad value growth, 8y.

-0.011 med 0.091 0.324

src Transfermarktvalue_cagr_8y

Form & fitness 10 features

Squad xG+xA per 90 (Understat top-5)

4 n/a

Squad club xG+xA per 90 (Understat top-5).

0.030 med 0.246 0.476

src Understatsquad_form_xgxa_per90

# squad players with Understat top-5 data

Coverage of top-5-league form data (0–1).

0 med 12.50 25

src Understatsquad_form_coverage

Squad fitness readiness (top-5)

4 n/a

Top-5-league minutes-based readiness.

0.020 med 0.513 0.814

src Oxford Football Forecasting modelsquad_fitness_readiness

Mean club minutes per game

4 n/a

Mean club minutes per game.

6.50 med 59.80 75.50

src API-Football (global club coverage)mean_minutes_per_game

Squad fitness (global)

GLOBAL (de-biased) fitness readiness.

0.158 med 0.707 0.886

src API-Football (global club coverage)squad_fitness_global

Global fitness coverage (fraction)

Coverage of global fitness data (0–1).

0.420 med 0.880 1

src API-Football (global club coverage)fitness_coverage_global

Squad club form (global, league-weighted)

GLOBAL (de-biased) squad club form.

0.052 med 0.211 0.410

src API-Football (global club coverage)squad_form_global

Squad club form (global, unweighted)

Global form before league normalisation.

0.072 med 0.148 0.250

src API-Football (global club coverage)squad_form_global_raw

Mean league-strength coefficient of squad clubs

Mean league-strength coefficient of squad's clubs.

0.502 med 1.41 1.80

src Oxford Football Forecasting modelmean_league_coef

Global club-form coverage (fraction)

Coverage of global form data (0–1).

0.420 med 0.880 1

src API-Football (global club coverage)form_coverage_global

Context (c) 6 features

GDP compound annual growth, 5y

GDP compound annual growth, 5y (World Bank).

-0.047 med 0.046 0.119

src World Bankgdp_cagr_5y

GDP compound annual growth, 10y

GDP compound annual growth, 10y.

-0.019 med 0.022 0.064

src World Bankgdp_cagr_10y

Squad familiarity / chemistry index

Shared-club chemistry index.

0 med 0.015 0.172

src Oxford Football Forecasting modelfamiliarity_index

# shared-club player pairs

Count of same-club player pairs.

0 med 5 56

src Oxford Football Forecasting modelshared_club_pairs

Minute-weighted familiarity index

Co-tenure-weighted chemistry.

0 med 0.026 0.434

src Oxford Football Forecasting modelfamiliarity_weighted_index

Climate-adaptation gap to host venues

1 n/a

Squad acclimatisation vs host max-temp (°C).

-18 med 1.70 9.60

src Open-Meteo & venue recordsclimate_gap

Decoupling (g) 2 features

z(NT form) - z(squad value): over/under-performance

Form residual vs market value.

-2.79 med 0.115 1.75

src Oxford Football Forecasting modelform_vs_value_gap

z(Elo) - z(squad value): history vs market

Elo residual vs market value.

-1.71 med 0.113 1.23

src Oxford Football Forecasting modelelo_vs_value_gap

Source · Oxford Football Forecasting model — the engineered feature panel (48 teams × 35 numeric features). Sparkline = a 12-bin histogram across the 48 nations.

League-strength normalisation

A player’s club form means more in a stronger league. Each of the 60 club leagues gets a relative-strength coefficient, estimated from market-implied club quality and shrunk toward a confederation prior for thin leagues. England tops it; the coefficient is what re-weights every player’s minutes and goals before they roll up into the squad-form signal. The top 12 are shown — query the explorer below for all 60.

League (country)	Clubs priced	Strength (z)	Form coef.	Example clubs
England	236	+2.21	1.8	Fulham, Brentford, Manchester City
Spain	111	+2.13	1.8	Rayo Vallecano, Real Betis, Sevilla
Germany	141	+1.84	1.8	SV Darmstadt 98, Eintracht Frankfurt, Bayer Leverkusen
Italy	82	+1.70	1.8	Empoli, Atalanta, Cittadella
France	95	+1.70	1.8	Marseille, Lille, Nice
Portugal	44	+1.14	1.49	Famalicao, Benfica, Sporting CP
Brazil	25	+1.03	1.433	Palmeiras, Santos, Atletico Paranaense
Netherlands	55	+0.74	1.297	Feyenoord, Twente, Ajax
Turkey	44	+0.49	1.186	Beşiktaş, Sivasspor, Bursaspor
Croatia	14	+0.34	1.124	Dinamo Zagreb, HNK Rijeka, HNK Hajduk Split
Russia	21	+0.27	1.098	Lokomotiv, Rubin, Baltika
Austria	10	+0.26	1.096	Grazer AK, Lask Linz, Red Bull Salzburg

Source · Oxford Football Forecasting model — league-strength estimates for all 60 leagues. Strength (z) is standardised; the form coefficient re-weights club form in the squad roll-up.

How the data becomes a forecast

The lineage, end to end: 20 raw sources are harmonised and engineered by 61 scripts into the processed panels, which feed the model ladder, which the simulator plays 1.1M times into the locked forecast this whole site reads.

Fig. V20 One direction, one path

Data lineage — raw to locked forecast

Raw 20 sources results · Elo · squads · API-Football · Transfermarkt · odds · World Bank · environment …
Engineer 61 scripts crosswalk · reproducible Elo · feature build · de-biasing · backtest panel
Processed ~35 tables derived_features · league_strength · squad_form_global · country_context …
Models 7 + ensemble Elo · Dixon-Coles · Bayesian · LightGBM → log-pool ensemble
Simulate 1.1M draws real 48-team bracket · full FIFA tiebreakers · MC-SE
Locked forecast this site one locked forecast → every page; SHA-256 stamped in the footer

There is exactly one path from data to page: every figure on this site is produced from the same locked forecast — which is why every number reconciles.

Source · Oxford Football Forecasting model

§ 04

The data explorer — live SQL in your browser

Query the full dataset live, entirely in your browser — no server involved, nothing uploaded. Run the worked examples, write your own SQL over all 12 tables, browse any table point-and-click, and export the result as CSV. Below it, the full schema is documented so the data layer is complete even with JavaScript off.

Try:

SQL runs in your browser — the engine loads on your first query.

Press Run (or ⌘/Ctrl + Enter) to execute the query above. Results — sortable, exportable — appear here.

Or browse a table:

What the explorer is (and what it is not)

The console runs a full SQL engine inside your browser — it loads on your first query, and nothing you type or compute ever leaves your machine. It is read-only: the published tables cannot be altered, and a stray UPDATE simply returns an error. Result tables are capped at 1,000 rows in the page for speed — the CSV export carries the full set. If the engine cannot load, the schema below still documents every table, column and type.

The schema — 12 tables

Every queryable table, its row count, what it holds, and (where documented) each column with its type and meaning. This is the no-JavaScript-complete reference for the explorer above.

forecast 48 rows · 46 cols The locked champion / stage probabilities + group-stage odds (the single source of truth).

Column	Type	Meaning
`rank`	int64	—
`code`	str	—
`cca3`	str	—
`team`	str	—
`p_R32`	float64	P(reach Round of 32 = clear group stage).
`se_R32`	float64	—
`p_R16`	float64	P(reach Round of 16).
`se_R16`	float64	—
`p_QF`	float64	P(reach quarter-final).
`se_QF`	float64	—
`p_SF`	float64	P(reach semi-final).
`se_SF`	float64	—
`p_Final`	float64	P(reach Final).
`se_Final`	float64	—
`reality_champ`	float64	P(champion), fixture-aware (the LOCKED forecast). = LOCKED p_Champion.
`reality_champ_se`	float64	Monte-Carlo standard error of champion prob.
`conformal_level`	str	—
`conformal_coverage_0p90`	float64	—
`conformal_mean_set_size`	float64	—
`group`	str	—
`confederation`	str	—
`power_rank`	int64	—
`power_champ`	float64	Fixture-FREE 'Power' champion prob (mean over constrained re-draws).
`power_champ_se`	float64	—
`power_se_mc`	float64	—
`draw_luck`	float64	actual_champ − power_champ; >0 = soft draw, <0 = tough draw.
`draw_luck_se`	float64	—
`bracket_half`	int64	—
`bracket_quadrant`	int64	—
`confed_color`	str	—
`is_host`	bool	—
`gs_p_pos1`	float64	—
`gs_p_pos2`	float64	—
`gs_p_pos3`	float64	—
`gs_p_pos4`	float64	—
`gs_p_win_group`	float64	—
`gs_p_top2`	float64	—
`gs_p_best_third`	float64	—
`gs_p_advance`	float64	—
`gs_exp_points`	float64	—
`gs_mean_gd`	float64	—
`gs_mean_gf`	float64	—
`gs_se_win_group`	float64	—
`gs_se_top2`	float64	—
`gs_se_best_third`	float64	—
`gs_se_advance`	float64	—

rankings 48 rows · 35 cols Power vs Reality ranks, champion odds and draw-luck per team.

Column	Type	Meaning
`power_rank`	int64	—
`team`	str	—
`group`	str	—
`confederation`	str	—
`actual_champ`	float64	—
`actual_champ_se`	float64	—
`power_champ`	float64	—
`power_se`	float64	—
`power_se_mc`	float64	—
`draw_luck`	float64	—
`draw_luck_se`	float64	—
`bracket_half`	int64	—
`bracket_quadrant`	int64	—
`bracket_half_if_runnerup`	int64	—
`bracket_quadrant_if_runnerup`	int64	—
`power_p_R32`	float64	—
`power_se_R32`	float64	—
`actual_p_R32`	float64	—
`power_p_R16`	float64	—
`power_se_R16`	float64	—
`actual_p_R16`	float64	—
`power_p_QF`	float64	—
`power_se_QF`	float64	—
`actual_p_QF`	float64	—
`power_p_SF`	float64	—
`power_se_SF`	float64	—
`actual_p_SF`	float64	—
`power_p_Final`	float64	—
`power_se_Final`	float64	—
`actual_p_Final`	float64	—
`power_p_Champion`	float64	—
`power_se_Champion`	float64	—
`actual_p_Champion`	float64	—
`code`	str	—
`reality_rank`	int64	—

teams 48 rows · 127 cols The full 48-team panel — forecast + every engineered feature + context, one wide row per nation.

Column	Type	Meaning
`rank`	int64	—
`code`	str	—
`cca3`	str	—
`team`	str	—
`p_R32`	float64	—
`se_R32`	float64	—
`p_R16`	float64	—
`se_R16`	float64	—
`p_QF`	float64	—
`se_QF`	float64	—
`p_SF`	float64	—
`se_SF`	float64	—
`p_Final`	float64	—
`se_Final`	float64	—
`reality_champ`	float64	—
`reality_champ_se`	float64	—
`conformal_level`	str	—
`conformal_coverage_0p90`	float64	—
`conformal_mean_set_size`	float64	—
`group`	str	—
`confederation`	str	—
`power_rank`	int64	—
`power_champ`	float64	—
`power_champ_se`	float64	—
`power_se_mc`	float64	—
`draw_luck`	float64	—
`draw_luck_se`	float64	—
`bracket_half`	int64	—
`bracket_quadrant`	int64	—
`confed_color`	str	—
`is_host`	bool	—
`gs_p_pos1`	float64	—
`gs_p_pos2`	float64	—
`gs_p_pos3`	float64	—
`gs_p_pos4`	float64	—
`gs_p_win_group`	float64	—
`gs_p_top2`	float64	—
`gs_p_best_third`	float64	—
`gs_p_advance`	float64	—
`gs_exp_points`	float64	—
`gs_mean_gd`	float64	—
`gs_mean_gf`	float64	—
`gs_se_win_group`	float64	—
`gs_se_top2`	float64	—
`gs_se_best_third`	float64	—
`gs_se_advance`	float64	—
`elo_now`	int64	—
`elo_change_5y`	int64	—
`elo_change_10y`	int64	—
`elo_slope_10y`	float64	—
`gdp_cagr_5y`	float64	—
`gdp_cagr_10y`	float64	—
`nt_last15_ppg`	float64	—
`nt_last15_gd_pg`	float64	—
`nt_last15_winrate`	float64	—
`nt_form_trend`	float64	—
`squad_form_xgxa_per90`	float64	—
`squad_form_coverage`	int64	—
`mean_age`	float64	—
`pct_peak_25_29`	float64	—
`n_over30`	int64	—
`n_under23`	int64	—
`mean_caps`	float64	—
`largest_club_bloc`	int64	—
`n_distinct_clubs`	int64	—
`squad_value_eur`	float64	—
`form_vs_value_gap`	float64	—
`elo_vs_value_gap`	float64	—
`value_cagr_8y`	float64	—
`familiarity_index`	float64	—
`shared_club_pairs`	int64	—
`familiarity_weighted_index`	float64	—
`climate_gap`	float64	—
`squad_fitness_readiness`	float64	—
`mean_minutes_per_game`	float64	—
`squad_fitness_global`	float64	—
`fitness_coverage_global`	float64	—
`squad_form_global`	float64	—
`squad_form_global_raw`	float64	—
`mean_league_coef`	float64	—
`form_coverage_global`	float64	—
`g_s_hat`	float64	—
`g_h_hat`	float64	—
`g_mean`	float64	—
`g_sd`	float64	—
`primary_language`	str	—
`n_languages`	int64	—
`shares_lang_with_host`	bool	—
`region`	str	—
`subregion`	str	—
`capital_lat`	float64	—
`capital_lon`	float64	—
`home_utc_offset`	int64	—
`gdp_usd`	float64	—
`gdp_per_capita`	float64	—
`population`	int64	—
`dist_nearest_venue_km`	int64	—
`mean_dist_played_km`	int64	—
`max_timezone_shift_h`	int64	—
`max_venue_altitude_m`	int64	—
`max_venue_heat_index_c`	float64	—
`coach`	str	—
`coach_nationality`	str	—
`appointed`	float64	—
`foreign_coach`	bool	—
`n_players`	int64	—
`mean_value_eur`	float64	—
`value_top3_share`	float64	—
`n_in_top5_league`	int64	—
`share_in_top5_league`	float64	—
`n_abroad_in_top5`	int64	—
`top5_npxg_plus_xa`	float64	—
`top5_minutes`	int64	—
`top5_league_breakdown`	str	—
`tm_value_coverage`	float64	—
`n_clustered_players`	int64	—
`n_club_blocs`	int64	—
`top_blocs`	str	—
`co_tenure_seasons`	int64	—
`pairs_with_shared_history`	int64	—
`diaspora_usa`	float64	—
`diaspora_can`	float64	—
`diaspora_mex`	float64	—
`diaspora_hosts_total`	float64	—
`diaspora_per_1000_pop`	float64	—
`squad_acclim_tmax`	float64	—
`host_tmax`	float64	—

players 1,245 rows · 22 cols All 1,245 squad players with their global club season (minutes, goals, league strength).

Column	Type	Meaning
`player_id`	str	—
`number`	int64	—
`position`	str	—
`player`	str	—
`dob`	str	—
`caps`	int64	—
`nt_goals`	int64	—
`club_squad`	str	—
`team`	str	—
`group`	str	—
`club_apifootball`	str	—
`league`	str	—
`league_country`	str	—
`club_minutes`	float64	—
`club_apps`	float64	—
`club_goals_season`	float64	—
`stats_matched`	bool	—
`league_strength_z`	float64	—
`league_log_q`	float64	—
`league_form_coef`	float64	—
`code`	str	—
`confederation`	str	—

coaches 48 rows · 8 cols Each nation’s coach, nationality and the foreign-coach flag.

Column	Type	Meaning
`team`	str	—
`coach`	str	—
`coach_nationality`	str	—
`appointed`	float64	—
`foreign_coach`	bool	—
`code`	str	—
`confederation`	str	—
`group`	str	—

matches 104 rows · 13 cols All 104 fixtures with venue, round and (where played) score.

Column	Type	Meaning
`match_no`	int64	—
`round`	int64	—
`date_utc`	str	—
`venue`	str	—
`home`	str	—
`away`	str	—
`group`	str	—
`home_score`	float64	—
`away_score`	float64	—
`winner`	float64	—
`stage`	str	—
`home_code`	str	—
`away_code`	str	—

features 48 rows · 39 cols The 36-column 2026 feature panel the models read (team key + 35 numeric).

Column	Type	Meaning
`team`	str	—
`elo_now`	int64	Current Elo rating (eloratings.net method, end-2026).
`elo_change_5y`	int64	Elo change over last 5 years.
`elo_change_10y`	int64	Elo change over last 10 years.
`elo_slope_10y`	float64	Linear Elo slope per year over 10y.
`gdp_cagr_5y`	float64	GDP compound annual growth, 5y (World Bank).
`gdp_cagr_10y`	float64	GDP compound annual growth, 10y.
`nt_last15_ppg`	float64	National-team points-per-game, last 15 matches.
`nt_last15_gd_pg`	float64	National-team goal-difference per game, last 15.
`nt_last15_winrate`	float64	National-team win rate, last 15.
`nt_form_trend`	float64	Recent NT form trend (slope).
`squad_form_xgxa_per90`	float64	Squad club xG+xA per 90 (Understat top-5).
`squad_form_coverage`	int64	Coverage of top-5-league form data (0–1).
`mean_age`	float64	Mean squad age.
`pct_peak_25_29`	float64	Share of squad aged 25–29 (peak).
`n_over30`	int64	Players over 30.
`n_under23`	int64	Players under 23.
`mean_caps`	float64	Mean international caps.
`largest_club_bloc`	int64	Largest same-club bloc size.
`n_distinct_clubs`	int64	Distinct clubs in squad.
`squad_value_eur`	float64	Total squad market value (Transfermarkt, EUR).
`form_vs_value_gap`	float64	Form residual vs market value.
`elo_vs_value_gap`	float64	Elo residual vs market value.
`value_cagr_8y`	float64	Squad value growth, 8y.
`familiarity_index`	float64	Shared-club chemistry index.
`shared_club_pairs`	int64	Count of same-club player pairs.
`familiarity_weighted_index`	float64	Co-tenure-weighted chemistry.
`climate_gap`	float64	Squad acclimatisation vs host max-temp (°C).
`squad_fitness_readiness`	float64	Top-5-league minutes-based readiness.
`mean_minutes_per_game`	float64	Mean club minutes per game.
`squad_fitness_global`	float64	GLOBAL (de-biased) fitness readiness.
`fitness_coverage_global`	float64	Coverage of global fitness data (0–1).
`squad_form_global`	float64	GLOBAL (de-biased) squad club form.
`squad_form_global_raw`	float64	Global form before league normalisation.
`mean_league_coef`	float64	Mean league-strength coefficient of squad's clubs.
`form_coverage_global`	float64	Coverage of global form data (0–1).
`code`	str	—
`confederation`	str	—
`group`	str	—

features_historical 128 rows · 17 cols The same features rebuilt leakage-safe for 128 past team-tournaments.

Column	Type	Meaning
`tournament`	str	—
`team`	str	—
`elo_asof`	float64	—
`elo_change_5y`	float64	—
`nt_last15_ppg`	float64	—
`nt_last15_gd_pg`	float64	—
`squad_form_xgxa_per90`	float64	—
`squad_fitness_readiness`	float64	—
`familiarity_index`	float64	—
`familiarity_weighted_index`	float64	—
`understat_coverage`	float64	—
`squad_fitness_global`	float64	—
`fitness_coverage_global`	float64	—
`squad_form_global`	float64	—
`squad_form_global_raw`	float64	—
`mean_league_coef`	float64	—
`form_coverage_global`	float64	—

model_results 2,128 rows · 14 cols Every out-of-sample match prediction (7 models × W/D/L) — the backtest evidence.

Column	Type	Meaning
`protocol`	str	—
`fold`	str	—
`model`	str	—
`match_id`	str	—
`home`	str	—
`away`	str	—
`pH`	float64	—
`pD`	float64	—
`pA`	float64	—
`actual`	str	—
`understat_tier`	str	—
`subgroup`	str	—
`p_over`	float64	—
`total_actual`	int64	—

collisions 28 rows · 8 cols The top-8 “earliest possible meeting round” pairs from the bracket draw.

Column	Type	Meaning
`teamA`	str	—
`teamB`	str	—
`groupA`	str	—
`groupB`	str	—
`earliest_round`	str	—
`detail`	str	—
`codeA`	str	—
`codeB`	str	—

league_strength 60 rows · 9 cols The 60-league relative-strength normalisation (England top → Qatar bottom).

Column	Type	Meaning
`league_country`	str	—
`n_priced`	int64	—
`raw_log_q`	float64	—
`shrunk_log_q`	float64	—
`strength_z`	float64	—
`prior_backed`	bool	—
`form_coef`	float64	—
`example_clubs`	str	—
`confederation_hint`	str	—

elo_current 336 rows · 3 cols Current world Elo for 336 teams — the reproducible rating spine.

Column	Type	Meaning
`team`	str	—
`elo`	float64	—
`rank`	int64	—

Keep reading

Models →How the seven models on this data become a forecast. Validation →The out-of-sample record: calibration, conformal coverage, the n = 3 limitation. Players →All 1,245 players in the same client-side SQL table.