bus.to.itLive across TorinoIllustration
Illustrative map. Measured paths and display fitting are separate; GPS gaps remain visible.
Neural predictions · Historical fallback · Observed service
Know when the bus actually arrives.
Live buses on the map, ETAs with a range when we're unsure, and a heads-up when a line leaves its route.
3 generationshistorical + two TCNs
4 methodsone matched comparison
20.7%lower error · October candidate*
* Latest offline candidate vs official realtime on 7 October: 4,349 forecasts / 1,594 arrivals. Arrival-balanced MAE; not a live service guarantee.
Illustrative app preview · times and vehicle counts below are examples
9:41
Where to?
Leaving soon
4
Falchera2 stops away · 3-5 min
3:429
Barriera LanzoArriving
Now68
Sassi5 stops away · 8-11 min
9:1513
GruScheduled
14:02Prediction research
Live historical model.
First TCN in shadow.
Latest TCN evaluated offline.
How it works
Where is the bus? What happens next.
We use public vehicle positions when available, identify progress along a compatible route pattern, and estimate the remaining journey from recorded history. Missing or repeated positions reduce what we can know.
- 01Match to the route
The GPS trace is matched to one official route pattern.
- 02Measure the road
Observed remaining travel times, with time-period estimates blended toward broader history when data is sparse.
- 03Show the range
A time window when we have a live bus. Scheduled when we don't.
68
Line 68 · SassiIllustrative route view
4
Falchera2 stops away · 3-5 min
3:429
Barriera LanzoArriving
Now68
Sassi5 stops away · 8-11 min
9:1513
GruScheduled
14:02What you get
Before you walk to the stop.
Interface illustrations; times and service examples are not live measurements.
Arrivals with a range
4
Falchera2 stops away · 3-5 min likely
3:4213
GruScheduled
14:02A window when there's a live bus. Scheduled when there isn't.
Route referencesBeta
2Earlier retained routes and latest published MaTO routes. Geometric differences still need stop-service validation.
Alerts
68
Line 68 divertedNear Sassi · 3 buses
Alerts when a line you follow diverts, and when it goes back on route.
Follows the street
Routes turn at corners. A diversion is easy to spot.
Coverage
lines in the 20 Oct published-route snapshot
Latest neural model
sarrival-balanced error · 7 Oct offline paired sample
The prediction modelsThree generations.
Learn from journeys
Read the recent
A newer model.
Three generations.
One clear status for each.
From recorded journey times to neural predictions: each generation builds on the historical estimator, which remains the fallback.
BusTO v1Historical baseline + fallback
Learn from journeys
already made.
Match a vehicle to its route, then look up the time earlier vehicles needed to reach your stop from that position. Similar time periods borrow from broader history when observations are sparse.
Route position → historical remaining time
BusTO v1.1Shadow experiment
Read the recent
movement pattern.
The first temporal convolutional network adds a learned correction using recent speed, progress, stopping and data quality. It remains a shadow experiment alongside passenger serving.
Historical estimate + first TCN correction
BusTO v2Live on the web
A newer model.
A better time contract.
The October TCN uses newer data, fixed one-minute history slots, shared training/serving features and known terminal waiting. It now serves web arrivals, with the historical estimator as fallback. The benchmark below is an offline evaluation.
Historical estimate + latest TCN correction
Measured on matching examplesFour methods.
Four methods.
The same arrival evidence.
20.7%lower arrival-balanced error than official
for the latest candidate on 7 October
for the latest candidate on 7 October
7 October 2026 · held-out arrival partition4,349 forecasts · 1,594 arrivals · fallback retained
Every arrival receives equal weight. Lower is better; error is in seconds.
Official realtimeCaptured public forecast
130.8sBusTO v1Historical estimator
127.0sBusTO v1.1First TCN + fallback
118.4sBusTO v2 candidateLatest TCN + fallback
103.7sSelected October replay: first TCN neural coverage 45.0%; latest 99.8%. Remaining cases use the historical fallback. August weights use their original preprocessor with the current compatible history wrapper; these are not recovered original live outputs.
Explore per-line categories, daily coverage and all metrics →Accuracy
ETAs are guesses. We say so.
Arrival uncertainty depends on distance, observation quality and available history. These illustrated labels explain the interface; the measured results above describe the selected evaluation samples.
9LiveNow
One stop awayLive bus on this trip, close enough for a tight ETA.
4Live3:42
A few stops outAbout 3-5 minutes. A window, not a fake exact time.
13Scheduled14:02
No live busNothing on the map for this trip. Timetable time, clearly marked.
68Betan/a
Possible deviationA live GPS signal can be flagged before it is confirmed. Counts may be incomplete or temporarily inaccurate.
Map
Check before you leave.
Explore live positions, recorded journeys and both route-reference views. The latest neural model serves web arrivals with a historical fallback.