Predicting bus arrivals when the feed reports positions but not delays: a reproducible three-layer method with an open reference implementation

Many public transport networks publish real-time vehicle positions but not delays, and many timetables give intermediate stop times interpolated at constant speed rather than observed. Bizkaibus, the interurban bus network of Biscay (Spain), has both properties: after the first hour of a trip, 85% of buses pass a stop more than two minutes before the published time. This paper specifies the arrival prediction method of the Bizkaibus+ app (Bide+) in enough detail to reimplement it in any language. The method uses timetable differences instead of absolute times; learns the actual travel time between consecutive stops by time band, as a median shrunk towards the all-day median of the segment; and predicts from three layers with explicit abstention rules: the GPS position projected onto the route, the last stop registered by the feed, and the timetable corrected trip by trip. It recalibrates itself daily, adopting only changes that reduce the error on held-out data. In 9.1 million predictions over 23 days, checked against the actual passage, the arrival shown is within ±2 minutes in 92–96% of cases when the bus is less than three minutes away and in 67% at ten to twenty minutes, against 23–25% for the published timetable. Without any live data, the corrected timetable alone doubles the accuracy of the published one. Files: the preprint (PDF) and an archive with the Python reference implementation, test vectors with 407 checks produced by the production engine, the public evaluator, the learned tables and the evaluation data. Code under the MIT licence; text and data under CC BY 4.0. Preprint, not peer reviewed. A longer technical report in Spanish is available at https://doi.org/10.5281/zenodo.24041572.Version 1.1 corrects version 1.0 after peer review by independent reviewers: it states the exceptions to temporal validation in the day-by-day evaluation, reports that a neural network using the same information improves on the method by 3–7% (the method is kept for auditability, not for accuracy), clarifies that the reference passage is not independent of layer 2, qualifies several figures (52% of all observed passages, 38% instead of 40%, the 97.8% of likely-departed trips), states which results cannot yet be recomputed from public data, and fixes the appendix numbering and several references. No result changes.The code and data archive (reference implementation, test vectors, evaluator, learned tables and evaluation data) is unchanged from version 1.0 and is available in that version of this record: https://doi.org/10.5281/zenodo.24285541

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-16
DOI
https://doi.org/10.5281/zenodo.24285540
Citations
1
Primary Topic
Transportation Planning and Optimization
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Predicting bus arrivals when the feed reports positions but not delays: a reproducible three-layer method with an open reference implementation

Yeray Garrido
1 citations
Zenodo (CERN European Organization for Nuclear Research)
Transportation Planning and Optimization
preprint

Predicting bus arrivals when the feed reports positions but not delays: a reproducible three-layer method with an open reference implementation

Yeray Garrido
preprint en
1 citations

Abstract

Many public transport networks publish real-time vehicle positions but not delays, and many timetables give intermediate stop times interpolated at constant speed rather than observed. Bizkaibus, the interurban bus network of Biscay (Spain), has both properties: after the first hour of a trip, 85% of buses pass a stop more than two minutes before the published time. This paper specifies the arrival prediction method of the Bizkaibus+ app (Bide+) in enough detail to reimplement it in any language. The method uses timetable differences instead of absolute times; learns the actual travel time between consecutive stops by time band, as a median shrunk towards the all-day median of the segment; and predicts from three layers with explicit abstention rules: the GPS position projected onto the route, the last stop registered by the feed, and the timetable corrected trip by trip. It recalibrates itself daily, adopting only changes that reduce the error on held-out data. In 9.1 million predictions over 23 days, checked against the actual passage, the arrival shown is within ±2 minutes in 92–96% of cases when the bus is less than three minutes away and in 67% at ten to twenty minutes, against 23–25% for the published timetable. Without any live data, the corrected timetable alone doubles the accuracy of the published one. Files: the preprint (PDF) and an archive with the Python reference implementation, test vectors with 407 checks produced by the production engine, the public evaluator, the learned tables and the evaluation data. Code under the MIT licence; text and data under CC BY 4.0. Preprint, not peer reviewed. A longer technical report in Spanish is available at https://doi.org/10.5281/zenodo.24041572.Version 1.1 corrects version 1.0 after peer review by independent reviewers: it states the exceptions to temporal validation in the day-by-day evaluation, reports that a neural network using the same information improves on the method by 3–7% (the method is kept for auditability, not for accuracy), clarifies that the reference passage is not independent of layer 2, qualifies several figures (52% of all observed passages, 38% instead of 40%, the 97.8% of likely-departed trips), states which results cannot yet be recomputed from public data, and fixes the appendix numbering and several references. No result changes.The code and data archive (reference implementation, test vectors, evaluator, learned tables and evaluation data) is unchanged from version 1.0 and is available in that version of this record: https://doi.org/10.5281/zenodo.24285541

Zenodo (CERN European Organization for Nuclear Research)
Sustainable cities and communities, Industry, innovation and infrastructure
Transportation Planning and Optimization
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon