⚽ World Cup AI Forecaster

Can a time-series "foundation model" predict the 2026 FIFA World Cup? A hands-on, honest case study in when modern AI helps a prediction problem — and when a 50-year-old algorithm does just as well.

---

I tried to forecast the 2026 World Cup with the AI everyone's hyping — a foundation model (Google's TimesFM) that predicts any series of numbers, zero training required. A football rating formula from the 1960s (Elo) beat it. Then one small change — feeding it a smoother signal — flipped it from worst to best. The real twist: do it right, and everything ties — the old formula, the cutting-edge AI, a trained meta-model, even the live betting market — all jammed against the same ceiling.

📝 Read the full story: docs/article.md · 🔬 Methodology: docs/METHODOLOGY.md · 📊 All results: docs/RESULTS.md

The one finding, in one chart

Same foundation model, two different inputs. Fed raw goal counts it's the worst model; fed a smooth signal (each team's Elo trajectory) the identical model becomes the best. The representation mattered more than the model.

What this is

A complete, reproducible forecaster for the 2026 World Cup that:

• Forecasts every match (win / draw / loss + expected score) from team strength,
• Simulates the whole tournament (groups → the official 48-team bracket → champion) via Monte-Carlo,
• Validates live against games already played and against a real prediction market (Polymarket),
• Pits a TimesFM foundation model against honest classical baselines (Elo, Dixon-Coles Poisson), a trained meta-model, and TabPFN (a tabular foundation model),
• Ships an MCP server so any LLM can use it as a forecasting agent.

Everything runs on free, no-key data (match history + fixtures from one public CSV; optional Polymarket odds from a public API).

Quickstart

git clone https://github.com/tpomerance/worldcup-ai-forecaster.git && cd worldcup-ai-forecaster
python -m venv .venv && source .venv/bin/activate

# install torch for your hardware first (https://pytorch.org), then:
pip install -r requirements.txt

# 1) pull data + show live group standings
python -m src.ingest

# 2) honest backtest: TimesFM vs Elo vs Poisson vs blend (add --timesfm)
python -m src.evaluate --timesfm

# 3) predict upcoming matches / a single match / the whole tournament
python predict.py --next 6
python predict.py --match "Spain vs Morocco"
python predict.py --simulate --sims 10000

# 4) full illustrated report (Markdown + PDF) and the figures
python -m src.report
python -m src.make_pdf

# tests
python tests/test_sanity.py

The forecaster's current call

	Model (simulation)	Live market (Polymarket)
Favorite	Argentina ~30%	France ~18%
Then	Spain · France · England	Argentina · Spain · England

(Numbers update as the tournament progresses — re-run python refresh.py.)

How well does it actually do?

Validated pre-kickoff on games already played, scored with RPS (lower = better):

	RPS	vs
Our model	0.168	—
Polymarket (the market)	0.152	sharper (markets are hard to beat)
Clueless (1/3 each)	~0.24	we're far better than chance

The model and the market agree on ~all match picks; its misses are genuine upsets that fooled the market too. See docs/RESULTS.md for the full tables and the surprise analysis.

Project structure

worldcup-ai-forecaster/
├── src/
│   ├── ingest.py            # data: results + fixtures + live standings (no API key)
│   ├── elo.py               # World-Football Elo ratings
│   ├── features.py          # time-decayed attack/defense + per-team series
│   ├── match_model.py       # Dixon-Coles / Poisson → W/D/L, scorelines
│   ├── forecast.py          # TimesFM forecasting (goals & Elo trajectory)
│   ├── baselines.py         # Elo, last-value Poisson, market-implied
│   ├── odds.py / polymarket*.py  # bookmaker + Polymarket (champion & per-match)
│   ├── blend.py / meta_model.py  # ensemble + trained stacker + TabPFN
│   ├── evaluate.py / validate_live.py   # backtests & live validation (RPS/Brier)
│   ├── simulate.py / bracket2026.py     # Monte-Carlo over the official bracket
│   ├── report.py / make_pdf.py / export_figures.py   # reporting & figures
│   └── ...
├── predict.py               # CLI: --match / --next / --simulate / --report
├── refresh.py               # daily refresh of data + report + figures
├── polymarket_mcp.py        # MCP server → LLM forecasting agent
├── agent_run.py             # run the agent over the MCP tools
├── tests/                   # sanity tests
├── docs/                    # article, methodology, results, MCP/agent guide
└── assets/                  # figures

The forecasting agent (MCP)

polymarket_mcp.py exposes the data + model + simulator as MCP tools, so any MCP-capable LLM client becomes a World Cup forecasting agent — pull live odds, get a model forecast, and simulate the bracket in plain English. Full wiring and an honest take on what the agent does (it relays the market, it doesn't out-predict it) in docs/MCP_AGENT.md.

Honest disclaimer

This is a study in forecasting methodology, not betting advice. The headline lesson is deliberately un-hyped: a foundation model is a tool with a shape — match it to the problem and it helps; otherwise a one-line baseline wins. The market remains the sharpest forecaster, and beating it is not the goal.

Acknowledgements

• Google Research — TimesFM
• martj42/international_results — match data
• Polymarket public API — prediction-market odds
• Prior Labs — TabPFN

License

MIT — see LICENSE.