Madison Metro ML
The Madison Metro API gives predicted arrival times. They are often wrong by 10-15 minutes. The underlying prediction is close to linear extrapolation: vehicle speed times remaining distance. It ignores route-level reliability patterns, stop-level variance, and weather.
Madison Metro ML learns the error in the API’s prediction, not the arrival time from scratch. If the API says 8 minutes and the model says +3, the corrected ETA is 11. The systematic biases (route, stop, hour, weather) are what a tree model picks up.
the model
47-feature XGBoost regressor. The features that matter: horizon terms (raw, squared, log, bucketed, long-horizon flag), cyclical sine/cosine encodings of hour, day, and month, rolling route-level and stop-level error histories with shrinkage toward the route mean for thin stops, live vehicle dynamics, and Madison weather from Open-Meteo. Feature engineering does most of the work; the model is a few hundred shallow trees.
honest uncertainty
Intervals come from Mondrian conformal prediction, calibrated to 90% coverage, stratified by route, day-type, and horizon bucket. A single global interval would be too wide on easy short-horizon weekday trips and too narrow on long-horizon Friday-night runs on unreliable routes. Conditioning calibration on those groups means each cell gets the interval earned from its own residuals, and the coverage guarantee holds per cell.
retraining behind a gate
The model retrains every night through GitHub Actions: pull recent outcomes, temporal split, fit, evaluate. It deploys only if the new model clears a hard gate: at least 2s MAE improvement over the model in production, on a test set large enough to trust. Most nights nothing deploys. That is the correct outcome.
what broke
The nightly pipeline was updated to save models in XGBoost’s native .ubj format, but the backend loader still looked for .pkl files. Model load failed silently; the backend’s fallback returned the raw API prediction, so the system appeared to work. I caught it when diagnostics showed zero ML predictions. Fix: track format in the model registry, try .ubj first. Lesson: fallback behavior hides failures. Monitoring needs to check that the model is actually being used, not just that the endpoint returns 200.
the map
React, DeckGL, MapLibre. 200+ live buses as WebGL points, not DOM markers, which is what keeps them moving at 60fps. Tap a stop for arriving buses with corrected ETAs and confidence bands. Mobile view is one full-screen map and a draggable sheet built for someone walking to a stop in gloves.
Live at madisonbuseta.com.