A data-driven forecast of the 2026 FIFA World Cup (USA · Canada · Mexico, 11 Jun – 19 Jul 2026): an interactive wall chart pre-filled by a statistical model, plus the dataset and pipeline behind every pick.
🔗 Live: https://wc26-predictions.boxd.sh — opens pre-filled with the model's predictions (all 104 matches + bracket). Bracket & champion are locked (pick: Spain). After the group stage the model was retrained on the actual results to predict the Round-of-32 scores and set the 4 knockout top scorers (Kane, Lautaro, Mbappé, Oyarzabal). Tap ℹ️ How it works, or the 💬 Ask the model chat, to see how any pick was made. Mobile-friendly.
Built and served live entirely on a boxd.sh VM — no deployment step: the VM is the server, so an edit is live the moment it's saved.
⚠️ Disclaimer: this is a fun, hobby-grade model — quasi data science, not a peer-reviewed forecast. It runs on free public data with plenty of simplifying assumptions and a snapshot taken days before kickoff. For entertainment, not betting advice.
index.html— the interactive wall chart: live group standings, an auto-advancing knockout bracket, saved in your browser, pre-filled with the model. Open it directly orpython3 -m http.server 8777.data/— a verified 4-year historical dataset for all 48 teams, plus enrichment (xG, context, lineups) and a current-form snapshot. The model's training base.model/— the prediction pipeline (Elo + Dixon-Coles + Monte-Carlo, blended with bookmaker odds and squad form). Full write-up in MODEL.md; the picks in SCORITO_PREDICTIONS.md.chat_server/— the FastAPI web app: serves the chart + the "Ask the model" transparency chat (a Claude tool-use agent grounded in the real model outputs). See chat_server/README.md.deploy/— systemd unit + env template + deploy notes for the live site.
The live site is served by the wc26-chat service (FastAPI: the chart + an "Ask the model" chat widget) from a clean webroot so .env/repo internals aren't web-exposed. The public chart is button-stripped (no Reset / Fill-with-model) so visitors can't wipe it. When predictions change: python3 model/fill_sheet.py → python3 model/build_public.py.
- Strength ratings — a World-Football Elo rating for every national team, learned from ~8,000 internationals (2018–2026), weighting World Cups above friendlies and bigger wins more.
- Scoreline model — a time-weighted Dixon-Coles (bivariate-Poisson) goals model turns those strengths into the probability of every exact scoreline per fixture (attack/defence per team + home advantage + a low-scoring-draw correction), fit by maximum likelihood.
- Current state — bounded nudges for injuries, suspensions, momentum and coach changes (researched per squad) — e.g. a team missing key forwards is marked down.
- Market calibration — blended with vig-free bookmaker odds (outright + group-winner), which corrected real model quirks (it underrated France/England, overrated Morocco/Japan).
- Tournament simulation — a 30,000-run Monte-Carlo of the full 48-team bracket (group tables → best-third placement → knockouts) gives each team's advancement and title probabilities.
- Scorito-optimal picks — every predicted scoreline is chosen to maximise expected Scorito points (
30·P(outcome) + 15·P(exact)), which is why so many picks are 1-0/0-1.
Out-of-sample it's ~26% better than a naïve baseline (RPS) with ~60% outcome accuracy, and beats both a plain Elo model and an "always 1-1" strategy. The group stage bore this out (59% outcomes, 1,335 Scorito pts, beating the naive baseline). After the groups the model is retrained on the actual results (model/predict_ko.py, simulate_ko.py, topscorers_ko.py) to predict the Round-of-32 scores + 4 phase top scorers; the bracket & champion (Spain) stay locked. Full write-up in MODEL.md; the picks in SCORITO_PREDICTIONS.md.
Window: 2022-06-07 → 2026-06-07 (4 years back from the brief date of 7 Jun 2026), played matches only.
Built from martj42/international_results (results.csv, goalscorers.csv, shootouts.csv), the standard open dataset of every men's international since 1872 — the spine. Layered on top: StatsBomb open-data (xG/shots) and api-football (match context). Raw pulls are cached for reproducibility. See SOURCES.md for every source + licence.
| file | rows | grain | built by |
|---|---|---|---|
matches.csv |
1,850 | one row per match (canonical) | build_dataset.py |
team_match_log.csv |
2,351 | one row per WC26 team per match (long / model-ready) | build_dataset.py |
team_summary.csv |
48 | per-team aggregate over the window | build_dataset.py |
deep_history.csv |
3,590 | one row per match, 8-year window (2018-06-07→2026-06-07), for rating stability | build_extras.py |
h2h.csv |
335 | head-to-head record for every pair of WC48 teams that met in the 4yr window | build_extras.py |
statsbomb_match_stats.csv |
173 | per-match xG/shots/cards (StatsBomb, 4 major tournaments) | fetch_statsbomb.py |
apifootball_context.csv |
654 | per-match stage/venue/referee/HT-ET-pen (api-football) | fetch_apifootball.py |
openfootball_match_extra.csv |
124 | attendance + starting XI + subs (WC 2022 + 2018, CC0) | parse_openfootball.py |
match_lineups.csv |
3,663 | long: one row per (match, team, player) with started/captain | parse_openfootball.py |
matches_enriched.csv |
1,850 | wide model-ready table: matches.csv + all enrichment, joined by match_id |
build_enriched.py |
MANIFEST.json |
- | build metadata + counts | build_dataset.py |
A match between two WC26 teams appears once in matches.csv and twice in team_match_log.csv (once from each team's perspective), which is why the long table has more rows.
python3 build_dataset.py # spine: matches / team_match_log / team_summary (stdlib)
python3 build_extras.py # deep_history.csv + h2h.csv (stdlib)
python3 fetch_statsbomb.py # statsbomb_match_stats.csv (downloads ~560MB events, cached)
python3 fetch_apifootball.py # apifootball_context.csv (needs .env key; free plan = 100/day)
python3 parse_openfootball.py # openfootball_match_extra.csv + match_lineups.csv (stdlib, CC0)
python3 build_enriched.py # matches_enriched.csv + coverage report (stdlib)fetch_statsbomb.py needs no key. fetch_apifootball.py reads APIFOOTBALL_KEY from a gitignored .env. Both cache raw pulls so re-runs are offline/free.
One row = one team's match. This is the "Netherlands played 49 games → 49 rows" view.
| column | meaning |
|---|---|
team / team_display / wc_group |
dataset name / wall-chart name / 2026 group (A-L) |
date / days_ago |
match date / days before 2026-06-07 |
opponent / opponent_is_wc26 |
opponent, and whether they're also a 2026 team |
venue / is_home_record |
home / away / neutral, and whether team was the listed home side |
gf / ga / goal_diff |
goals for / against / difference, from this team's view |
result / points |
W/D/L and 3/1/0 |
went_to_shootout / shootout_won |
knockout shootout flags |
tournament / competition_type / is_competitive |
raw competition / coarse bucket / friendly-vs-not |
city / country / neutral |
where it was played |
scorers |
this team's goals, e.g. Memphis Depay 23' (pen); Cody Gakpo 67' |
match_id / notes |
join key to matches.csv / reserved (see Enrichment) |
match_id, date, days_ago, home_team, away_team, home_score, away_score, total_goals, goal_difference, result (H/D/A), winner, loser, went_to_shootout, shootout_winner, tournament, competition_type, is_competitive, city, country, neutral, home_is_host, wc26_home, wc26_away, both_wc26, home_scorers, away_scorers, notes
competition_type buckets the raw tournament into: World Cup, WC Qualifier, Continental Cup, Continental Qualifier, Nations League, Friendly, Other. The raw string is always preserved.
Everything in matches.csv plus, joined on match_id (blank where no source has it):
StatsBomb columns (prefix home_/away_, 173 matches): sb_stage, home_xg/away_xg, home_shots/away_shots, home_sot/away_sot (shots on target), home_corners/away_corners, home_fouls/away_fouls, home_yellow/away_yellow, home_red/away_red, home_passes/away_passes. xG and shots exclude penalty shootouts (period 5) so they reflect 0–120′ play; goal counts reconcile 173/173 to the real scoreline (incl. own goals).
api-football context columns (654 matches): af_competition, af_round (stage + matchday, e.g. Group Stage - 1, Final), venue_name (stadium), venue_city, referee, ht_home/ht_away (halftime), et_home/et_away (extra-time score, 34 matches), pen_home/pen_away (shootout score, 26 matches).
openfootball columns (63 WC 2022 matches): attendance, home_xi/away_xi (the 11 starters, ;-joined), has_lineups flag. Full per-player detail (starters + subs + captain, WC 2022 and 2018) lives in match_lineups.csv.
deep_history.csv mirrors the match-level columns over an 8-year window with an in_4y_window flag (TRUE subset == the 1,850 canonical matches). h2h.csv: team_a, team_b, played, a_wins, draws, b_wins, a_goals, b_goals, last_meeting, last_score, last_tournament (record always from team_a's perspective; pairs sorted alphabetically).
| field group | matches | coverage | source |
|---|---|---|---|
| xG / shots / cards (StatsBomb) | 173 | 9.4% | WC22, AFCON23, Copa24, Euro24 |
| stage / venue / referee / HT (api-football) | 654 | 35.4% | comps in seasons 2022–2024 |
| extra-time score | 34 | 1.8% | — |
| shootout score | 26 | 1.4% | — |
| attendance (openfootball) | 63 | 3.4% | WC 2022 only |
| starting XI / lineups (openfootball) | 63 | 3.4% | WC 2022 (+ WC 2018 in deep table) |
| either enrichment layer | 661 | 35.7% | — |
How in-shape each of the 48 teams is, four days before kickoff. Built in two layers: derived hard signals from the match data (free, 100% covered) + researched qualitative info (coach changes, injuries, news, momentum), every claim cited.
| file | rows | grain | built by |
|---|---|---|---|
team_state.csv |
48 | per-team snapshot: derived form + coach/qualification/momentum/shape verdict | build_part2.py |
team_injuries.csv |
77 | one row per injury / doubt / suspension — 100% cited | build_part2.py |
team_news.csv |
144 | one row per dated news item — 100% cited | build_part2.py |
team_key_players.csv |
232 | one row per key-player note | build_part2.py |
team_state_form.csv |
48 | derived-only: last5/10 form, 2026 warm-up results, streaks, rest days | build_team_state.py |
Per-team research is cached as data/part2_raw/<team>.json (coach, qualification path,
injuries, key players, warm-up read, momentum, shape verdict, news — all with source URLs).
Coverage: head coach 48/48, qualification 48/48, shape verdict 48/48, 23/48 teams had a
coach change in the last year, injuries found for 36/48 (the rest honest blanks). Build:
python3 build_team_state.py && python3 build_part2.py. Full 26-man rosters are the one
remaining Part 2 gap (we have 3–5 key players/team). See SCHEMA_PART2.md.
Output is checked, not assumed.
Spine (build_dataset.py):
- date range within window; 0 rows with NA scores; 0 rows without a WC26 team
- coverage re-counted independently from raw vs the fresh martj42 master: 1,850/1,850 played WC48 internationals captured, 0 duplicate
match_id. Cache is byte-identical to the current martj42 master. - per-team counts sit in a plausible 37-68 range (Netherlands 49; small FAs like Haiti/Curaçao/New Zealand at 37; CONCACAF/Asian sides higher from Gold/Asian Cups)
- scorelines spot-checked: 2022 WC Final Argentina 3-3 France (so Argentina) ✓; Euro 2024 Final Spain 2-1 England ✓; Copa América 2024 Final Argentina 1-0 Colombia ✓
Extras (build_extras.py): deep_history 4yr subset == 1,850; h2h integrity a_wins+draws+b_wins == played for all 335 pairs; spot-checks Argentina 4-1 Brazil (2025-03-25), Spain 5-4 France ✓
StatsBomb (fetch_statsbomb.py): 173/173 matches have non-empty stats for both sides; derived goals reconcile 173/173 to the real scoreline after excluding shootout penalties (period 5) and crediting own goals; xG spot-checks match known values (WC22 final 2.76–2.27). Penalty-shootout shots were initially (and incorrectly) inflating xG — caught and fixed.
api-football (fetch_apifootball.py): stage/venue/referee/HT verified against known finals (WC22 Lusail, ref Marciniak, HT 2-0, pens 4-2; Euro24 Olympiastadion Berlin; Copa24 Hard Rock).
- xG / shots only for the 173 big-tournament matches (StatsBomb open data). FBref, which would extend xG to qualifiers/friendlies, is Cloudflare-blocked from this VM's datacenter IP (403).
- api-football context stops at season 2024 on the free plan, so the 2025–2026 WC qualifying cycle (CONMEBOL/UEFA/AFC/CONCACAF/OFC) and 2025 tournaments have no stage/venue/referee. A paid plan would unlock them.
- Attendance and lineups cover only WC 2022 (openfootball, CC0); WC 2018 lineups are in the deep table. Other competitions have neither — openfootball is WC-finals-only, and the 2026 qualifiers aren't in it. Wikipedia/Wikidata would be the source for the rest.
- Injuries / managers are Part 2 (see SCHEMA_PART2.md), not yet collected. The
notescolumn is reserved.
- Part 2 — recent team developments: schema designed in SCHEMA_PART2.md (form, squad, injuries, manager/tactics, qualification path, momentum). Not yet collected.
- Prediction model — built (see MODEL.md): Elo + time-weighted Dixon-Coles goals model + Monte-Carlo of the 2026 bracket, tuned to maximise Scorito points. Out-of-sample skill ~26% RPS over climatology, 60% outcome accuracy. Pre-tournament champion pick Spain (blended model+market); retrained after the group stage to predict the Round-of-32 scores + 4 top scorers; bracket & champion (Spain) locked. Full submission in SCORITO_PREDICTIONS.md; scoring in SCORITO_RULES.md.
Full table in SOURCES.md. In short: match spine = martj42/international_results (open, attribution); xG = StatsBomb open data (CC BY-NC-SA, attribution required); context = api-football (commercial, free tier); BBC_WC_26_WALL_CHART.pdf is BBC copyright (reference only, do not redistribute). API keys live in a gitignored .env.
🍺 Licence rider: if this model wins your Scorito pool, you owe the developer a beer.