Matched outcomes
The official-paired October candidate sample requires exact trip, line and stop, realtime forecasts and issue times within 45 seconds. Each official observation is used once.
bus.to.itRead the methodology ↓Arrival prediction · Evidence update
Better estimates start
with a fair comparison.
Historical travel times, a small temporal neural model, and comparisons with the official realtime field. Every result belongs to its own dates, sample and deployment stage.
What would each method have predicted?
Four methods.
One matched October sample.
20.7%lower arrival-balanced error: latest candidate versus official
in this selected offline sample
in this selected offline sample
7 October held-out arrival partition · new frozen replay4,349 forecasts · 1,594 arrivals · baseline fallback retained
Each of the 1,594 arrivals receives equal weight. Lower is better.
The two neural systems include the same historical fallback. August weights use their own scaler/vocabulary and the current legacy-compatible sequence wrapper on retained October features. This is a counterfactual model replay, not recovered original live outputs. No weights were retrained or tuned.
Support changed as well as prediction quality.
The August model produced neural predictions for 1,958/4,349 forecasts; the latest model for 4,342/4,349. Unknown inputs remain in the graph through fallback. One selected test day, strict-label selection and public-feed-derived outcomes limit generalisation.
What happened in September?
The strongest shared realtime-field transition was between the captured polls at 09:23:21 and 09:27:21 on 14 September, Europe/Rome. Captured realtime/all rows changed from 0/95 to 86/95 on Line 4, 0/47 to 39/44 on Line 9 and 0/40 to 18/34 on Line 68. This identifies a feed transition, not an internal release timestamp or proven estimator update.
15 September morning is a separate intermittent input-quality episode: at 09:00–09:59, median last-received report age increased to 30.2/49.4/64.6 seconds on the three lines. We cannot infer original GPS-fix age from receipt time. The October graph is not a reconstruction of how October-trained weights would have performed before training.
How to read the evidence
Dates, definitions and coverage stay visible.
Observed passage proxy
Next-stop transitions are refined with GPS and quality filters. Shared feed timing and uncertain historical route validity remain limitations.
Stage of deployment
The October TCN 2 serves web/PWA arrivals with historical fallback. The August TCN remains a shadow experiment. The scores on this page are frozen offline or historical evaluations, not a new prospective service measurement.
Prepared from frozen local reports on 2026-10-20. Publishing these reports does not change their original evaluation dates or constitute a new live measurement.