Artificial Wasteland · 50 days of the Swiss rail record

Where the Lateness Goes

A timetable is not a statement about how fast a train can go. It is that plus a supplement, spread along the route so a train that falls behind can catch up. The railway literature sets norms for how much of it there should be and argues about where to put it. Nobody has published how much of it a national network actually gives back. Here is 6,485,058 measured legs of one, and the answer has a ceiling in it.

That is the whole finding, and the slider is the argument. Move it to one minute and the lateness is essentially gone by the destination. Move it to fifteen and almost none of it is. The Swiss timetable is a filter: it removes small lateness completely and passes large lateness through nearly untouched.

What was measured, and on what

Switzerland publishes, every day, a file with one row per train per stop carrying both the scheduled time and the realised time. This reads the publisher's whole rolling window as it stood on 2026-09-13: 2026-07-25 to 2026-09-12, 50 days, 8,902,734 rail rows from 53 operators, 737,701 usable train-days and 6,485,058 station-to-station legs where both ends carry a measured time rather than a forecast.

The publisher is explicit that most of what is in the file is not a measurement, and the distinction is a field. Its own cookbook gives the enumeration:

UNBEKANNT (no forecast or actual times for this and all preceding stops) · PROGNOSE (arrival forecast) · GESCHAETZT (calculated actual arrival time) · REAL (effective actual arrival time)

Nothing here uses a value whose status is not REAL. That is also why this page does not simply count minutes late: a delay figure built from forecasts would be measuring the forecaster.

The number a passenger wants

Take every train that left its first station d whole minutes late and ask how late it was when it reached its last. No model, no estimate of how fast the train could have gone, no parameter to choose: a difference of two published timestamps, twice.

A train one minute late arrives 10 seconds late: 83% of the lateness is gone. Two minutes late becomes 33 seconds. Five minutes late becomes 2 min 45 s, so barely half of it is recovered. Ten minutes late becomes 7 min 47 s, and fifteen minutes late becomes 14 min 37 s, which is 97% of what it started with.

Why it has a ceiling

The mechanism is visible one leg at a time. For every station-to-station leg, bin the trains by how late they left the first station and take the median change in lateness across the leg. That is how much the leg gave back.

A leg gives back 19 seconds to a train that is barely late, rises to 48 seconds at 3 minutes, and then falls away. It never exceeds about 48 seconds however late the train is. The median leg class gives back 39 seconds at its best, and 1,260 of 5,824 leg classes give back nothing at all.

So the arithmetic is forced. Recovery is bounded per leg, which means the lateness a journey can undo is bounded by how many legs it has, not by how much help it needs. A train ten minutes down would need something like fifteen consecutive maximally generous legs, and journeys do not have them. Past roughly eight minutes the railway stops being able to help and simply carries the lateness to the destination.

Where the slack is kept

Two places, and both are measurable separately. Standing, not running. Of the scheduled time a train spends stationary at a platform, 60% is above the first percentile of what it is observed to need; of scheduled running time, 20% is. Dwell is where a timetable hides its money.

In front of the junction. Switzerland runs a clock-face timetable, in which trains must reach the interchange nodes just before the hour and the half hour so that connections can be made. That should push slack into the last leg before a hub, and it does: the mean best absorption of a leg ending at Zürich HB, Bern, Basel SBB, Olten, Luzern, Lausanne or Genève is 87.0 seconds against 45.4 seconds everywhere else.

What it costs

Slack is not free: it is journey time that the passenger spends and the train does not need. Comparing a route's scheduled running and dwell time against the first percentile of what those same legs are observed to take gives a lower bound on the share of the clock that is supplement rather than travel.

Long-distance categories only, by trip count. The forty-route table the pre-registration fixed is in the payload below and is dominated by suburban shuttles; this is a display choice and changes no number in it.

Seven predictions, written down first

The estimators, the exclusions and seven falsifiable predictions were committed before analyse.mjs existed, in research/where-the-lateness-goes/PREREGISTRATION.md. The verifier asks git for both commit times and fails if they are the wrong way round. 4 held and 3 broke.

The interesting break is the last one. The prediction was that slack would not buy punctuality across segments, because planners put slack where trains are already unreliable, which should cancel or reverse the association. It does not: the correlation is positive (Spearman 0.193). That is not evidence that adding slack would improve punctuality, and this page does not claim it is. It is observational, the allocation is not random, and a positive sign under a confound that should have pushed the other way is a fact that wants an experiment nobody can run from outside.

The crosscheck, and where it lands

SBB's own 2025 annual report gives train punctuality as 94.1 per cent, defining a train as punctual when it arrives with less than three minutes' delay and counting cancelled trains as unpunctual. The Federal Office of Transport, over subsidised regional traffic for the same year, gives 95.37 per cent of train runs. Computed here from the open file, over every rail operator and every stop, dropping cancellations, the same threshold gives 93.32%; counting every cancelled stop that had a scheduled arrival as unpunctual instead gives 90.36%.

The bracket contains neither published figure exactly and it should not: SBB reports its own trains, the Federal Office reports regional lines at designated measuring points, and this counts all 53 operators at every stop. The point of the comparison is that an independent route through public data lands two or three points from the official number rather than somewhere else entirely, which is the most this crosscheck can honestly buy.

The legs that cannot have happened

Asking how long a train took between two stops is a question with no negative answers, so any negative answer is a defect. There are 24,337 of them among 6,485,058 legs, concentrated in 193 leg classes, and 11 of the largest 25 share one cause: the departure recorded at the first stop is really the departure from the second. In 8 of them the two stops carry the same departure timestamp, to the second, in most of their pairs.

These are small halts on regional lines. Nothing here says a train was mis-run; it says the record cannot state when it left the first of two adjacent stops, because it reuses one event for both. Those traversals are dropped from every estimator on this page, and the count is published rather than swept up.

Source

Ist-Daten (actual data), Open Data Platform Mobility Switzerland, Open Data Platform Mobility Switzerland (ODMCH) / opentransportdata.swiss. https://data.opentransportdata.swiss/dataset/ist-daten-v2. Terms of use: https://opentransportdata.swiss/en/terms-of-use/, which state that no registration or payment is required, that the data may be processed, analysed and published, and that opentransportdata.swiss must be cited as the source. Each day's file is recorded in data/sources.json with its URL, byte length and sha256 as received.