MultiForecast

MultiForecast is a dual-mode football forecaster: a Premier League season simulator and a World Cup 2026 tournament predictor in one Streamlit app. Both products share a Monte Carlo and statistical-modeling foundation, while keeping league-table logic and group-plus-knockout tournament logic cleanly separated.

Architecture

app.py                         # Streamlit entry point with two tabs
premier_league/
  dashboard.py                 # Premier League tab
  refresh.py                   # Daily PL refresh workflow entry
  run.py                       # PL simulation runner
  engine/                      # Premier League engine modules
  viz/                         # PL analysis notebooks
worldcup/
  ui.py                        # World Cup tab
  data.py, model.py            # WC loaders and match models
  tiebreakers.py, bracket.py   # FIFA ranking and knockout logic
  simulate.py                  # WC tournament Monte Carlo
data/
  *.csv, *.json                # Premier League data
  wc/                          # World Cup data layer
cache/                         # App and simulation caches
.github/workflows/
  daily-refresh.yml            # Automated PL refresh

Shared Concepts

Both products use rating-based match probabilities, scoreline or result sampling, Monte Carlo aggregation, cache-aware outputs, and a daily GitHub Action refresh pattern. They diverge where football formats diverge: the Premier League product simulates a 20-team league table, while the World Cup product simulates groups, FIFA tiebreakers, third-place qualification, and a knockout bracket.

Run Locally

streamlit run app.py

Premier League refresh and pipeline:

python -m premier_league.refresh
python -m premier_league.run

World Cup simulator smoke run:

python -m worldcup.simulate

Automated Daily Refresh

The GitHub Action Daily Refresh (PL + World Cup) runs python refresh_all.py every day at 10:00 UTC and commits updated tracked data/ outputs. Premier League and World Cup live fetch steps are non-fatal, so a flaky upstream source prints a warning and the job continues with the last-good data where possible.

Product Docs

• Premier League predictor
• World Cup 2026 predictor

Modeling Roadmap

The modeling path is Elo first, then Poisson/Skellam scoreline models, then Dixon-Coles calibration, and finally richer ML features once stable data and validation loops are in place.

Changelog

• v2.0: World Cup 2026 predictor and monorepo restructure.
• v1.x: Premier League Monte Carlo dashboard.