How it works.

A train that has been steadily two minutes late for five stops and a train that has just picked up an unexplained delay both get the same kind of estimate on a board: one time, stated with the same confidence.

The range is what separates them. With a four minute connection at the other end, a two minute range says you are fine and a twelve minute range says you might not be. A single number cannot tell you that.

The formal name is a prediction interval: a low bound and a high bound, chosen so the true answer falls between them a stated share of the time. This one aims for 80%. That is a claim anyone can check against what actually happened, which is the whole point of the accuracy page.

By asking for the same train on the same date in 2020, 2024, 2025 and 2026. The times all differ, and the 2020 version has three fewer stops and a different scheduled arrival at Cork, which is the 2020 timetable on 2020 infrastructure. It is genuine history, back to at least 2007.

That single check is what makes a missed collection run recoverable. It does not apply to the operator's expected arrival, which exists only live and can never be backfilled, which is why a poller has been running every five minutes since August.

28,706 gzipped responses covering 1,087 train codes across 34 dates, about half a million stop level records. They were collected once, at two requests per second, overnight, backing off whenever the server complained.

A real arrival almost never lands on exactly the scheduled second. Across the whole archive, 2.92% of arrivals are identical to the scheduled time.

Far more of the flagged lines' records were not machine captured: 47% against 1.5% elsewhere, and the non-automatic records are where echoes live. Among machine captured records the flagged lines echo less; among hand-entered records they echo more. The line name was standing in for how the data was captured.

A comparison that reverses when you split it has a name: Simpson's paradox. Three quarters of the suspect records sit on lines the documentation never mentions, so a line filter would have thrown away good data and kept bad.

An exact match is suspicious, not proof. About 2.4% of the most trustworthy records genuinely do land on the scheduled second. Part of the reason is that every time in the feed is rounded to six seconds, so there are only ten possible second values in a minute and coincidences are commoner than they look.

So nothing is deleted at ingestion. The flag is carried through, what to exclude is decided when evaluating, and the numbers are reported both ways. A result that agrees with the documentation is exactly the one nobody checks carefully enough, and this one was wrong.

With lopsided penalties. An ordinary model is trained to be wrong by as little as possible in both directions, which gives a middle. To learn the 90th percentile, being too low is punished nine times as heavily as being too high, so the model settles where 90% of outcomes fall below. This is called quantile loss.

The models themselves are gradient boosted trees: many small decision trees, each correcting what the ones before it got wrong.

Delay accumulated upstream today, stops and minutes remaining, time of day, day of week, and a route proxy. All twelve inputs exist for a train that started running yesterday. The strongest by far is the delay at the previous stop, because lateness carries forward along a journey.

A model that had learned "A218 runs two minutes down" would have nothing to say about a new service, which would arrive as an unknown category, or about the same service after a renumbering.

Nothing forces them to agree, so occasionally the 10th percentile comes out above the 90th. That is called quantile crossing. The three values are sorted at prediction time, which is provably no worse than leaving them crossed.

It is the baseline being compared against, so feeding it in would make the comparison meaningless.

Three rules, each there because ignoring it produces a flattering wrong number.

• Matched events only. A result counts only where both the model and the operator produced a prediction for the same train, at the same station, at the same moment. Comparing all of one against all of the other compares populations, not methods.
• One result per event, not per poll. A single arrival is polled about eighteen times as the train approaches. Counting those as eighteen independent results inflates the sample size roughly eighteenfold and makes noise look significant.
• Matched precision. The operator publishes to the minute and the model to the second, so the model's answer is rounded to the minute before either is scored.

It was not looked at through any model change, and the analysis was written down and committed before it was opened. Once opened, every later decision would be made knowing the result. A test set that has been looked at is no longer a test set, so any future held-out check needs new data.

The read pattern is "give me one whole day" and the write pattern is append only. Neither needs one, and a database that stays up bills by the hour whether or not anyone asks it anything.

Measured on 23 August 2026, before the site was added, about ten cents a month, almost all of it storage requests.

An accuracy page built on a handful of the author's own test queries would be worthless, and almost nobody visits this site. So a scheduled job asks the live service a spread of questions every five minutes, one per lead band, sampled across whatever is in service rather than favouring any route.

Those are real predictions, made before the outcome existed. They are simply not anyone's, and the accuracy page says so. Predictions requested by this site's boards carry a different source tag, so the two can always be told apart.

On 2 September 2026 three Sligo line trains each lost 75 to 89 minutes in a single stretch. The service had been asked sixteen times across those journeys, and at every one of those moments the train was two to seven minutes late. Nothing in "how late is it now, how far is left to go" can see that coming.

A range wide enough to catch it would be wrong about ordinary days to hedge against rare ones, so the ranges are calibrated for ordinary lateness.

That has not been established. The retraining rule fired on those three groups, and the ranges were not widened to hide it: coverage is published per line, recomputed every night, rather than blended into one number.

Four of them were introduced by the author, after the system was working, while writing up the others. The lesson drawn was not "be more careful". It was to attach a denominator and a check to every published number.

Almost certainly some. Eleven is a count of the ones that were caught, and each was caught only because the number it touched had a denominator attached and a check that asked what it should have been. That is the design, and it is the argument for publishing where the model loses rather than only where it wins.