gamblistics-lab/sport-statistics

Football Match Data

Match-level, team-level, and player-level football data collected from various sources, intended for research on match outcome and predictive-market modeling (final score, half-time score, corners, over/under goals, time of first goal, etc.).

Data is provided as JSON Lines (.jsonl) — one JSON object per line.

Files & Row Shapes

Row counts reflect the current snapshot; the dataset may grow over time.

matches.jsonl — 1,140 rows

One row per match.

• id — match UUID (join key for match_stats, match_player_stats)
• kickOff — Unix epoch seconds (int)
• homeTeam, awayTeam — team names (denormalized)
• homeTeamId, awayTeamId — join to teams.id
• homeScore, awayScore — full-time score (strings, cast to int)
• homeScore1st, awayScore1st — 1st-half score (may be string or int)
• homeScore2nd, awayScore2nd — 2nd-half score
• homeScore90, awayScore90 — score at 90 min (before extra time, if any)
• status — integer enum; 3 = full-time (finished)
• leagueId → leagues.id; tournamentId → tournaments.id; stageId → tournament_stages.id
• roundName — e.g. "Round 24"

match_stats.jsonl — 760 rows

Team-level stats per match, split into three periods.

• id — same UUID as the matching matches.id
• matchId — duplicate of id
• overall, first_half, second_half — each is {metric: {home, away}}
• Metrics observed: xg, possession, shots_total, shots_on_target, shots_off_target, shots_blocked, corners, passes_total, passes_accurate, yellow_cards, offsides, free_kicks, throw_ins, fouls, gk_saves.

match_lineup.jsonl — 47,837 rows

One row per (match, player) — the raw upstream form (one doc per match with an embedded player array) has been exploded.

• id — "<matchId>_<playerId>" (composite)
• matchId, playerId, teamId
• homeFormation, awayFormation — match-level, repeated on every row
• side — "home" or "away"
• isStarter, isCaptain, isGoalkeeper — booleans
• playerType — 1 = player, 2 = coach, 3 = goalkeeper
• jerseyNumber, position, formationSlot, pitchRow
• playerName, playerCountry, playerCountryId

Note: minuteOn/minuteOff are absent for full-90 starters and must be inferred as (0, 90 + stoppage) downstream.

match_player_stats.jsonl — 442 rows

Per-player stats per match. Coverage is heavily biased toward top leagues and matches from 2024-25 season onward — many matches will not have a row here. Model design must account for MNAR (missing-not-at-random) gaps.

• id — same UUID as the matching matches.id
• players — array of {playerId, teamId, side, stats: {METRIC: value}}

Stat keys are SCREAMING_SNAKE_CASE — e.g. EXPECTED_GOALS, SHOTS_TOTAL, PASSES_ACCURATE, PASSES_OPEN_PLAY_ACCURACY, DUELS_EFFICIENCY, FS_RATING. Up to ~103 keys per player; every player carries the full dictionary with 0s for non-events (e.g. PENALTY_SHOOTOUT_* in a regular league match). Do not treat these 0s as observed absence of the event — drop constant-zero columns or emit a metric_observed mask.

Dimension tables

• players.jsonl (1,296) — id, name, country, countryId, primaryPosition, playerType, dateOfBirth
• teams.jsonl (23) — id, name, countryId, countryName, stadium, city, capacity
• leagues.jsonl (1) — id, name, countryId, sportId — permanent competition (e.g. "Premier League")
• tournaments.jsonl (115) — id, leagueId, season — one season of a league
• tournament_stages.jsonl (115) — id, tournamentId, leagueId, typeId, match_count
• sports.jsonl (1) — id, name

Join Keys

• matches.id == match_stats.id == match_player_stats.id for the same match.
• match_lineup.id == "<matchId>_<playerId>"; use matchId to join to matches.id.
• matches.homeTeamId / awayTeamId → teams.id.
• match_lineup.playerId / match_player_stats.players[].playerId → players.id.
• matches.leagueId → leagues.id; matches.tournamentId → tournaments.id.
• players.countryId / teams.countryId → country UUID (country dimension not included in this release).

Known Data Quality Caveats

• match_stats and match_player_stats do not cover every match — they are gated by upstream availability; expect ~two-thirds of matches to lack team stats and most to lack per-player stats.
• Score fields in matches are strings in some rows and ints in others — cast to nullable int before use.
• Timestamps are Unix epoch seconds (not milliseconds).
• Metric naming is inconsistent between match_stats (snake_case) and match_player_stats.players[].stats (SCREAMING_SNAKE_CASE). Normalize in your loader.
• The proprietary source format is reverse-engineered and can shift silently — sanity-check field distributions before every training run.

Suggested Uses

• Predicting final score as home/away Poisson rates (λ_home, λ_away)
• 1st-half score and half-time result markets
• Over/Under 1.5 & 2.5 goals
• Total corners
• Time of first goal
• Player-level xG / shot / rating modeling

Any classification/regression targets used for betting markets should be followed by calibration (Platt or isotonic) on a held-out set — uncalibrated probabilities are misleading for expected-value calculations.

Source & Licensing

Data was collected from various sources. This release is provided under CC-BY-NC-4.0 — non-commercial, with attribution — for research and educational purposes only. If you build on this dataset, please credit the uploader and note the original source.

No personal data beyond publicly displayed player and coach names, nationalities, and dates of birth is included.

Citation

If you use this dataset, please cite:

@dataset{gamblistics_sport_statistics,
  title  = {Gamblistics Sport Statistics},
  author = {Nguyen Thuc Tuyen},
  year   = {2026},
  url    = {https://huggingface.co/datasets/gamblistics-lab/sport-statistics/}
}