Track record
How accurate our predictions have been, against MBTA's own, on predictions the model never saw.
Tested on 2026-11-03T06:00+00:00 to 2026-11-03T11:58+00:00 (UTC), 1,500,189 predictions: ours were within MBTA's accuracy windows 92.8% of the time against 87.6% for MBTA's, and off by 125 s on average against 170 s. The range we give held 79.7% of the time (it aims for 80%).
| Mode | Predictions | Ours within window | MBTA's within window | Our mean error | MBTA's mean error |
|---|---|---|---|---|---|
| Subway and Green Line | 69,031 | 95.3% | 90.6% | 109 s | 164 s |
| Bus | 1,415,749 | 92.5% | 87.3% | 127 s | 171 s |
| Commuter Rail | 13,739 | 98.6% | 95.2% | 84 s | 111 s |
| Ferry | 1,670 | 98.1% | 96.2% | 52 s | 78 s |
Model blend-20261101T062950Z (experimental), trained on 2026-10-24T22:00+00:00 to 2026-10-31T22:00+00:00. We only show our time where it has beaten MBTA's; the model is retrained every 8 hours and must beat the one serving to replace it.