sergiopesch/worldcup-2026-predictions

World Cup 2026 — Multi-Model Match Predictions

Pre-match win/draw/loss predictions for the FIFA World Cup 2026 fixtures, produced by four independent prediction models and served live by a public demo app. The opening two fixtures predate three model pipelines; those missing pre-match forecasts remain null rather than being reconstructed after the fact.

This is a faithful export of data generated by a Salesforce-backed prediction pipeline — no synthetic rows.

The four models

model_version	Name	Approach
baseline-v3.1	The Analyst	Explainable weighted model (human-written rules)
elo-v1	The Historian	Elo-style team-rating model
ml-xgboost-v1	The Machine	External gradient-boosted ML model (the companion model repo)
einstein-studio-v1	The Oracle	Salesforce Einstein Studio, native in Data 360

Files

predictions.csv — one row per (match × model)

column	description
match_id	stable fixture id
kickoff_utc	kickoff time (UTC)
home, away	team codes
model_version, model_name	which model produced the row
home_win_prob, draw_prob, away_win_prob	outcome probabilities (%)
pick, pick_side	the model's called outcome
confidence	margin-based confidence (%)

teams.csv — 48 qualified teams

FIFA ranking, fifaPoints, group, confederation, isHost, etc.

matches.csv — all 104 tournament fixtures

Stable tournament order, canonical stage, kickoff, teams, status, final score, venue, and city. Stage cardinalities are 72/16/8/4/2/1/1.

Example

from datasets import load_dataset
ds = load_dataset("sergiopesch/worldcup-2026-predictions", "predictions")
print(ds["train"][0])

import pandas as pd
df = pd.read_csv("predictions.csv")
# Where do the four models disagree most?
spread = df.groupby("match_id")["pick"].nunique()
print(spread[spread > 1])

Intended uses

• Compare modelling philosophies (rules vs ratings vs ML vs AutoML) on identical fixtures.
• Teach probability calibration / model-agreement analysis.
• Benchmark your own model against four baselines.

Limitations

• All group and knockout fixtures are included through the final.
• Predictions are pre-match and reflect each model at export time.
• Team codes follow FIFA three-letter conventions.

Provenance

Generated by a World Cup 2026 prediction app (Salesforce worldcup org → public Node/React web app). Historical training data for The Machine comes from martj42/international_results.

License

Released CC-BY-4.0. Attribution appreciated. Built as a learning/demo project; not affiliated with or endorsed by FIFA.