Trenalia

Data validation

From 2026-10-19 to 2026-11-02 · computed 2026-11-02 20:16 UTC

How accurate Trenalia and its sources are, measured against what happened afterwards: every night, what the site showed is compared with the real arrival, measured by Trenalia from station boards and from the GPS position recorded throughout the journey. It is fully automatic: nothing is entered by hand.

8.7 minmean error of Trenalia's delay (official: 7.9 min)

38 %samples within ±3 min of the real arrival delay

33journeys with an observed arrival (3,499 samples)

68.8 %network alerts confirmed (240 checked)

1 · What happened at the end of the journey

Trenalia measures on its own when each train arrived: from the destination station's arrivals board (ADIF or RadarDeTrenes) when available, otherwise from the continuously recorded GPS position (entering within 700 m of the station, or an estimate from the last position and speed if it vanishes within 3 km). That arrival is compared here with Renfe's last published delay. Positive bias = Renfe reports more delay than observed.

	Samples	Mean difference (min)	Agree within ±2 min	Renfe bias (min)	p90
All observed arrivals	33	4.8	45.5 %	3.1	14
Destination arrivals board	27	5.5	40.7 %	4.2	15.4
GPS: entering the station	4	2	75 %	-2	5.2
GPS: estimated when signal lost nearby	2	2	50 %	-2	2.8
Boards as the train passes each station	2,078	2.8	70.8 %	-0.3	6

2 · Delay shown vs. observed arrival

The toughest test: what the site showed at each moment is compared with the arrival measured by Trenalia (board or GPS), not with Renfe's figure. “Official” is Renfe's published delay; “Trenalia” the projected one. Gain = how much Trenalia reduces the official error.

	Samples	Official error (min)	Trenalia error (min)	±3 min official	±3 min Trenalia	Bias (min)	Gain
Total	3,499	7.9	8.7	47.2 %	38 %	0.7	-9.9 %
0–15 min before arrival	426	5.3	6	50.5 %	42.7 %	5.1	-13.6 %
15–30 min before arrival	448	5.1	6.2	48.4 %	42.2 %	5.4	-22.3 %
30–60 min before arrival	761	4.6	5.9	53.2 %	43 %	4.3	-27.4 %
60–120 min before arrival	1,011	5.5	6.9	45 %	32.2 %	2.5	-25.4 %
2–4 h before arrival	853	16.5	16	42.1 %	36 %	-9.1	+3.2 %

By service

	Samples	Official error (min)	Trenalia error (min)	±3 min official	±3 min Trenalia	Bias (min)	Gain
Media Distancia and regional	1,484	12.5	12.9	38.6 %	27.4 %	-3.9	-2.7 %
Euromed and Intercity	863	4	4.2	47.9 %	43.5 %	2	-6.9 %
AVE	608	2.3	4.2	74.3 %	62.3 %	3.6	-86.3 %
Alvia	227	12.3	13.6	0 %	0 %	13.6	-10.8 %
Avant	200	2.5	2.6	86.5 %	82.5 %	-0.5	-2.2 %
Avlo	117	7.9	12.2	34.2 %	5.1 %	12.2	-53.7 %

By day

	Samples	Official error (min)	Trenalia error (min)	±3 min official	±3 min Trenalia	Bias (min)	Gain
2026-11-01	128	10.4	12.1	10.2 %	3.1 %	12.1	-16.3 %
2026-11-02	3,371	7.8	8.5	48.6 %	39.4 %	0.3	-9.6 %

3 · Delay shown vs. Renfe's last figure

The same, taking as final the last delay Renfe published at the end of the journey (available for almost every journey, but not an independent measure). It shows whether Trenalia anticipates what Renfe will end up reporting.

	Samples	Official error (min)	Trenalia error (min)	±3 min official	±3 min Trenalia	Bias (min)	Gain
Total	9,376	7.1	8.2	50.9 %	45.6 %	0.2	-16.2 %
0–15 min before arrival	1,004	1.1	4.2	89.7 %	72.4 %	3.8	-284.8 %
15–30 min before arrival	1,086	2.5	4.7	74.5 %	62.1 %	3.4	-83.2 %
30–60 min before arrival	1,927	3.7	5.3	54 %	46.4 %	3	-43.3 %
60–120 min before arrival	2,826	7.4	7.3	40.4 %	39.1 %	1	+0.7 %
2–4 h before arrival	2,533	13.6	14.6	34.9 %	34.6 %	-5.6	-7 %

By service

	Samples	Official error (min)	Trenalia error (min)	±3 min official	±3 min Trenalia	Bias (min)	Gain
Media Distancia and regional	5,424	9	11	46.8 %	38.4 %	0.1	-22.6 %
AVE	2,155	3.6	3.6	65.6 %	65.6 %	0.5	-1 %
Euromed and Intercity	863	4.4	4.4	55.2 %	52.5 %	-1.3	+0.8 %
Alvia	548	7.7	7.9	26.6 %	27.6 %	3.7	-2.1 %
Avant	251	3.1	3.6	68.1 %	54.2 %	2.7	-14.6 %
Avlo	135	9.7	6	24.4 %	29.6 %	-6	+37.7 %

By day

	Samples	Official error (min)	Trenalia error (min)	±3 min official	±3 min Trenalia	Bias (min)	Gain
2026-10-31	578	11.3	12.7	55.9 %	44.6 %	1	-12.2 %
2026-11-01	4,480	5.8	6.8	47.8 %	46.7 %	1.4	-18.5 %
2026-11-02	4,318	7.9	9.1	53.6 %	44.6 %	-1.1	-15.2 %

4 · Learned models

Trained every night on the history and evaluated on the last week, which is held out from training. They are only switched on if they clearly beat the baseline.

Delay correction (ETA)

needs more data faltan datos · trained 2026-11-01 04:28 UTC · — of — needed

False alarm filter for closing alerts

needs more data faltan alertas etiquetadas · trained 2026-11-01 04:28 UTC · — of — needed

5 · Train closing alerts

There is no official data to compare them with, so we label what happened afterwards. If the gap kept closing, the alert was consistent with the data; if the train ahead was stopped at a station, it was probably a false alarm (normal operations).

Candidates evaluated	949
Alerts shown	880
Still closing 2 min later	11.8 %
Likely false alarm (train ahead at a station)	0.9 %

What happened afterwards

Automatic check 3 h after each alert shown: whether both trains finished their journey normally, whether Renfe or Adif published an incident in the area, and whether there were network alerts nearby. “No signs” is not official confirmation: official accident data is published by the CIAF.

No signs	327
Incident in the area	215
Mentioning collision or derailment	0
Signs to review	170
No data	38
Reviewed by hand	0

6 · Network alerts

A segment or stopped-train alert counts as confirmed if any train involved gained 3 min or more of delay 10 to 30 min later; a sudden delay jump, if it did not revert (it was not a data glitch).

	Alerts	Checked	Confirmed
Segment with several trains losing time	9	9	77.8 %
Train stopped outside a station	213	193	62.2 %
Sudden delay jump	39	38	100 %
Total	261	240	68.8 %

7 · Data quality

Days with trajectories	3
Position samples per day	75,720
Positions snapped to the track	99.1 %
Estimated precision (median)	144 m · p90 748 m

Free software (GNU GPL v3)