BahnMining - Pünktlichkeit ist eine Zier (David Kriesel)

Summary

David Kriesel presents an analysis of Deutsche Bahn (DB) punctuality and cancellations using a large dataset he collected from public railway data interfaces. The talk is framed as an effort to understand how the rail network works—not simply to attack the railway—and as a demonstration of how to build and check a data project.

Dataset and approach

• Kriesel collected roughly 25 million stop records, beginning on January 8. The data covers long-distance services across Germany and local services at long-distance stations.
• Each record represents a train’s stop, with information such as scheduled and actual timings, delays, and cancellations. This allows analysis by station, train, route, and time period.
• DB defines a stop as punctual if the train is less than six minutes late. Using that measure, DB reported roughly 75% punctual stops in 2018.
• Kriesel’s dataset broadly reproduces DB’s published punctuality figures, with small differences. He presents this as an external check that the data is reasonably reliable, while noting that it is not exact or complete.
• He also checks the data internally—for example, by visualizing stop activity and cancellations over time. The plots show normal daily patterns as well as major disruptions, such as a storm that brought much of long-distance traffic to a halt.
• The dataset has limitations: it does not cover all local-transport stations, and replacement services may not always appear. Kriesel cautions that the results should be treated as an external analysis rather than an official DB accounting.

Main findings

• Punctuality varies substantially by region. Long-distance stops in much of eastern Germany performed better in the analysis, while North Rhine-Westphalia performed poorly. Cologne was around 66% punctual and Bonn around 59%.
• Some stations appear to add delays to trains, while others help reduce them. Frankfurt Central Station had the largest estimated contribution to delays—about 93,000 minutes during the measurement period. Bremen, Berlin Central Station, and Berlin-Spandau appeared among the stations that reduced delays. Kriesel notes that his station measure includes the surrounding track segments, not just the station itself.
• Punctuality differs by train type. Intercity trains were around 76% punctual under DB’s definition, while EuroCity trains were below 70%. International routes may be affected by delays originating outside Germany.
• Cancellations are a significant part of the picture. Kriesel estimates cancellation rates of a little over 3% for Intercity trains, about 2% for EuroCity trains, and over 5% for ECE trains. He also observes a seasonal increase in cancellations, particularly for Intercity services in summer.
• Longer journeys tend to become less punctual. In the data, punctuality falls as a train’s journey progresses; after around 500 minutes of travel, fewer than 60% of stops are punctual.
• A delay of 40 minutes or more is a point after which trains in the analysis were more likely to remain delayed than to recover.
• Kriesel estimates that 12.4% of stops were either more than 20 minutes late or cancelled. Under the ticket rules he discusses, such a disruption may remove a train-specific restriction on a saver ticket. He identifies particular combinations of train, station, and weekday with high historical rates of that outcome, but warns that past patterns are not guarantees for a future trip.

Why cancellations matter

• DB’s punctuality calculation excludes cancelled stops from its denominator. Kriesel argues that this can make the punctuality rate look better than passengers’ experience: a train that does not run cannot count as late under that calculation.
• As an alternative, he proposes counting all planned stops and then measuring how many were actually punctual. Using this approach, he estimates punctuality at about 72.5%, rather than the roughly 75% figure produced by excluding cancellations.
• He also points out a possible incentive problem: cancelling a delayed service can improve the punctuality statistic if the cancelled stops disappear from the calculation.
• His analysis suggests that cancellations may occur more often near the beginning and end of some train journeys. He discusses this as a possible operational response to accumulated delays, while acknowledging that DB may make such decisions to limit wider disruption.
• His criticism is primarily about how the statistics represent cancellations: the negative effect on affected passengers is not reflected in the punctuality figure.

Guidance for data projects

Kriesel outlines several lessons from building the project:

• Plan data collection around the source’s interfaces. DB’s timetable and change information are accessed separately, and the interfaces provide only limited time windows, so the data must be collected repeatedly.
• Minimize requests and avoid burdening the service. A naïve approach querying every station every few minutes could generate an enormous volume of traffic. Kriesel instead limited collection to roughly 350 long-distance stations and queried at most hourly, reducing the request volume substantially.
• Build monitoring for the collection process. He lost data when his download system and database failed. His technical lesson is to monitor the downloader independently, ideally on a separate server.
• Check legal and usage conditions. He recommends reading the terms, seeking legal advice if necessary, and obtaining written permission when appropriate. He says DB granted him written permission to collect and present the data.
• Avoid selective reporting. Do not choose only the time periods that support a preferred conclusion.
• Verify the data. Compare your results with published figures where possible, then check whether the data behaves plausibly using visualizations and other internal checks.
• Treat conclusions cautiously. A close match with a source’s published figures supports confidence in the dataset, but does not prove that every record is correct or complete.

Broader message

Kriesel argues for evidence-based public discussion. Rather than relying on outrage or sensational claims, he encourages people to investigate contentious topics, show their methods, and revise their views when the data suggests something unexpected. He also stresses that the railway has useful functions and that the analysis is not intended as simple “Bahn-bashing.”

Speakers and sources featured

• David Kriesel — presenter and creator of the analysis.
• Conference introducer/announcer — introduces the talk.
• Session moderator — thanks the speaker and manages the Q&A.
• Audience members — ask questions during the Q&A; they are not identified by name.
• Deutsche Bahn — source of the timetable and service-change data, as well as the published punctuality figures and documentation discussed in the talk.
• Spiegel — cited by Kriesel as having published a separate analysis with broadly similar cancellation estimates.