Algorithm: scikit-learn HistGradientBoostingClassifier (gradient-boosted trees)
Training data: sergiopesch/wc2026-internationals — 499 historical internationals between the 48 finalists
5-fold CV accuracy: 0.545 (vs 0.455 majority-class baseline)
Live demo + API: sergiopesch/wc2026-match-predictor Space

Features

Engineered to be symmetric / rating-based so they generalise monotonically (raw rank/points levels were deliberately excluded — they let the trees memorise this friendly-heavy sample and invert on unseen mismatches):

feature	meaning
`rank_diff` = away_rank − home_rank	>0 favours home
`pts_diff` = home_pts − away_pts	>0 favours home
`home_is_host`	home side is a 2026 host (USA/Mexico/Canada)

Usage

python

import joblib, pandas as pd
from huggingface_hub import hf_hub_download

model = joblib.load(hf_hub_download("sergiopesch/wc2026-match-predictor", "model.joblib"))
X = pd.DataFrame([{"rank_diff": 11 - 5, "pts_diff": 1776 - 1694, "home_is_host": 0}])
print(dict(zip(model.classes_, model.predict_proba(X)[0])))

Or call the hosted API:

bash

curl -X POST https://sergiopesch-wc2026-match-predictor.hf.space/predict \
  -H "Content-Type: application/json" \
  -d '{"home_rank":5,"home_pts":1776,"away_rank":11,"away_pts":1694,"home_is_host":0}'

How it's used

Registered in Salesforce Einstein Studio as a Bring-Your-Own-Model, scoring live 2026 fixtures from a Data Cloud feature pipeline — one of four prediction engines in a World Cup demo (alongside a transparent rules model, an Elo rating system, and a native Einstein Studio model).

Limitations

Three coarse features capture broad strength gaps, not form, injuries, or tactics. Educational/demo use; predictions are illustrative. Licence: CC0-1.0.

sergiopesch/wc2026-match-predictor

WC2026 Match Predictor — ml-xgboost-v1

Features

Usage

How it's used

Limitations

sergiopesch/
wc2026-match-predictor

WC2026 Match Predictor — `ml-xgboost-v1`