2026-08-21 to 2026-09-10 · measured on 1,495,332 logged prediction moments · updated 2026-09-11
Each point is one service day: the share of MetraX forecasts landing within one displayed minute of the real arrival, minus the same day’s share for the naive baseline, measured on identical rows. Above the dashed line, MetraX beat simply carrying the feed’s last reading forward unchanged. The series begins the day the current model went live; a day the engine has not scored leaves a gap rather than being drawn through.
| horizon out | n | Metra's native ETA within-1-min | MetraX app within-1-min |
|---|---|---|---|
| [0,5) min | 133658 | 99.4% | 99.0% |
| [5,15) min | 144339 | 77.2% | 80.9% |
| [15,30) min | 27905 | 41.0% | 46.6% |
| [30,60) min | 4238 | 23.9% | 27.2% |
Same moments, same stops, paired rows only (series since Aug 21, 21 day(s)). The MetraX column replays the shipped predictor under its live anchor policy — when a train is minutes away, MetraX intentionally carries the feed's own number, so the columns converge at short horizons; the model's added value appears from 15 minutes out, where a rider is deciding when to leave. GPS anchoring is not included in this replay, so the MetraX column slightly understates the live app.
| element | within-1-min lift | error change |
|---|---|---|
| Route history | +7.5 pts | −9.7% |
| Delay momentum | +4.1 pts | −15.6% |
| Signal freshness | +0.2 pts | −1.6% |
| Full blend (shipped) | +6.9 pts | −16.8% |
Each element measured against the naive baseline on identical rows (paired), trained only on data prior to each scored day.
Every metric on this page is computed from MetraX's own logged observations replayed against realized arrivals — no hand-picked examples, no survivorship. "Within 1 minute" means the displayed arrival minute was at most one minute off. The shipped-model series begins 2026-08-21.