Touchline

What's inside

• Fixtures and forecasts for 22 competitions: score probabilities, expected goals, over 2.5 and both teams to score.
• Match spotlight with the most likely score, an AI-written preview for upcoming fixtures, and a question box that answers from the forecast facts shown on the page.
• Weekly AI recaps summarising the last seven days of results for each competition, using only the saved scores.
• Match simulator with formations, a replay, events, pause and skip, reproducible seeds, and a 10,000-run comparison.
• Club and player profiles: squads, form, managers, stadiums, honours and career histories.
• Honest backtests: outcome and exact-score hit rates, log loss and Brier score, with the forecast cutoff shown.

How it works

• Forecasts come from a Rust model in crates/predictor, built from saved match data. Every forecast is dated, so the cutoff is always visible.
• The site is built from saved data. tools/build_website.cjs compiles the dashboard and packages it as static pages with on-demand JSON. No server is needed to view it.
• Results refresh daily. A scheduled workflow pulls open results from openfootball, moves the forecast date to today, rebuilds, and republishes. It only commits when the data changes.
• AI features are cached. Fixture previews (data/ai/previews.json) and weekly recaps (data/ai/recaps.json) are generated once per item by the DeepSeek API. The question box runs as a Vercel serverless function (api/match-question.js), with a per-visitor limit.

Build and run

Requirements: stable Rust and Node.js 20 or later.

node tools/build_website.cjs            # build the site into dist/
node tools/serve_dashboard.cjs dist/index.html

Open http://127.0.0.1:8765/. Previews need a DEEPSEEK_API_KEY in the environment. Without one, the site builds and simply shows no previews.

Run the checks:

cargo test --locked -p predictor
node --test crates/predictor/web/*.test.cjs

Configuration

Variable	Where	Purpose
DEEPSEEK_API_KEY	GitHub Actions secret	AI match previews during the scheduled build
DEEPSEEK_MODEL	Optional	Model name. Defaults to deepseek-chat.

Data and licences

• Results: openfootball, public domain.
• Team and player profiles and images: bundled from the sources recorded in data/. Each image keeps its own credit and licence, which the site shows.
• Code: dual-licensed under MIT or Apache-2.0.

Forecasts are statistical estimates for analysis and entertainment. They are not betting advice.

Project structure

crates/
  predictor/        forecasting model, backtests and the dashboard generator
  football_ai/      match simulation engine
  football_sandbox/ tactical sandbox (development tool)
data/               saved results, catalog, profiles and AI previews
tools/              build, refresh and publishing scripts
.github/workflows/  scheduled refresh and publishing
docs/               website publishing notes and screenshots

Licence

MIT or Apache-2.0, at your option.