Note
🪦 Retired — no longer maintained (June 2026). This was a fun side project: real-time Madison Metro bus tracking with ML-powered arrival predictions. The backend, database, and nightly training pipeline have been wound down, so the live app and predictions aren't running anymore.
The code stays public as-is — fork it, read it, learn from it, whatever you like. It was fun while it lasted. Thanks to anyone who gave it a shot. 🚌
— Nils
Real-time bus tracking with ML-powered arrival predictions for Madison, WI. Built because the official Metro app doesn't cut it.
Live: madisonbuseta.com (retired)
- Live map of all 29 Madison Metro routes with real-time vehicle positions (DeckGL + MapLibre, 60fps with 200+ markers)
- ML arrival predictions -- XGBoost regression corrects the Metro API's estimates using 47 features (temporal, route history, stop history, weather, vehicle speed). Currently 35% more accurate than trusting the API raw
- Confidence intervals via Mondrian conformal prediction -- statistically guaranteed 90% coverage, stratified by route x day-type x horizon
- Bus bunching detection -- two buses within 500m for 2+ consecutive polls, highlighted on the map with amber route-snapped overlays
- Mobile PWA -- responsive bottom-sheet UI with geolocation-based nearby stops, installable as a home screen app
- Analytics dashboard -- MAE trends, hourly bias, route reliability scores, bunching frequency, feature importance
+------------------+
| Madison Metro |
| REST API |
+--------+---------+
|
+--------------------------+---------------------------+
| |
+------------v-----------+ +-------------v----------+
| Collector (Railway) | | Flask Backend |
| 60s vehicle polls | | (Railway) |
| arrival detector +----------+ | /vehicles, /predict |
| bunch detector | | | /all-patterns, etc. |
+------------------------+ | +-------------+----------+
| |
+---------v---------+ |
| PostgreSQL | |
| (Railway) <---------------------+
| 30-day rolling |
+---------+---------+
|
+---------v---------+
| Nightly Training |
| (GitHub Actions) |
| XGBoost + gates |
| Conformal cal. |
+-------------------+
|
+---------v---------+
| React Frontend |
| (Vercel CDN) |
| Static routes |
| DeckGL map |
+-------------------+
The collector polls every 60 seconds, tracking all active buses. When a bus comes within 30m of a stop (haversine), it logs the actual arrival time. This gets matched against the API's prediction to compute error_seconds -- the training target.
| Category | Features |
|---|---|
| Horizon | horizon_min, horizon_squared, horizon_log, horizon_bucket, is_long_horizon |
| Temporal | hour_sin/cos, day_sin/cos, month_sin/cos, is_weekend, is_rush_hour, is_holiday, is_morning_rush, is_evening_rush |
| Route | route_frequency, route_encoded, predicted_minutes, route_avg_error, route_error_std, hr_route_error, route_horizon_error, route_horizon_std, dow_avg_error |
| Stop | stop_avg_error, stop_error_std (shrinkage estimators for stops with <50 samples) |
| Weather | temp_celsius, is_cold, is_hot, precipitation_mm, snow_mm, is_raining, is_snowing, is_precipitating, wind_speed, is_windy, visibility_km, low_visibility, is_severe_weather |
| Vehicle | avg_speed, speed_stddev, speed_variability, is_stopped, is_slow, is_moving_fast, has_velocity_data |
The nightly pipeline won't deploy a model that isn't better:
- Minimum 2s MAE improvement over current model
- MAE must be under 90s
- At least 1,000 test samples
- No regression allowed on any metric
- Temporal split: train on days 1-5, test on days 6-7
After training, a Mondrian conformal prediction layer calibrates uncertainty bounds per stratum (route x day-type x horizon bucket). Global coverage target: 90%. The calibration artifact is committed alongside the model.
Desktop and mobile are separate UIs sharing the same MapView component.
Desktop: Three-tab layout (Map / Analytics / System) with a context panel for route drilldowns, stop predictions, and trip planning.
Mobile: Full-screen map with a draggable bottom sheet. Geolocation-based nearby stops, tap-to-predict, bus tracking bar. Installable as a PWA.
Route shapes are served from a static JSON snapshot (~200KB gzipped) -- zero API calls to draw the map.
| Layer | Tech |
|---|---|
| Frontend | React 19, TypeScript, Vite 7, Tailwind 4, DeckGL 9.2, MapLibre 5.13, Recharts, Framer Motion |
| Backend | Flask, SQLAlchemy, Python 3.11 |
| ML | XGBoost 1.7.6, scikit-learn, Mondrian conformal prediction |
| Database | PostgreSQL (Railway, 30-day retention) |
| Infra | Railway (backend + collector + DB), Vercel (frontend + CDN), GitHub Actions (nightly training) |
| Domain | madisonbuseta.com (Namecheap + Vercel) |
# Backend
cd backend
pip install -r requirements.txt
cp .env.example .env # add MADISON_METRO_API_KEY
flask run --port=5000
# Frontend
cd frontend
npm install
npm run dev # http://localhost:5173The frontend works read-only without an API key (route map renders from static data). Live vehicles and ML predictions require MADISON_METRO_API_KEY from Madison Metro's developer portal.
Madison Metro's API has a daily request cap. If the app detects the API is out of requests (0 buses during service hours), it shows a banner explaining the situation and linking to an email template for requesting a higher quota from the city.
This project is retired and no longer actively maintained, but the code is MIT-licensed and free to fork. Questions or curiosity: nilsmatteson@icloud.com.
MIT License