Ten years of solar-storm forecasts have moved the arrival-time error by 1.2 hours

Across 1,702 logged predictions of when a coronal mass ejection would reach Earth, the average miss is 13.2 hours. The scoreboard's first five years averaged 14.0; the five and a half years since, 12.8.

Noodle News data desk · reading time ~3 min

13.2h

Mean absolute error across every scoreboard prediction that matched an observed arrival, May 2013 to December 2023

17.4h

Standard deviation of those errors — the spread, not the average, is what a planner has to absorb

1.2h

Fall in mean absolute error between the 2013–18 predictions and those since, which the authors call small and possibly not significant

What the record shows

Eleven models, one ten-to-fourteen-hour band

Since May 2013 NASA's Community Coordinated Modeling Center has run an Arrival Time Scoreboard, on which forecasting groups log, before the event, when they expect a coronal mass ejection to reach Earth. A 2024 review of every entry to the end of 2023 counted 2,617 submissions from 39 models, of which 1,702 corresponded to an arrival that was actually observed. The mean absolute error was 13.2 hours, the median 10.1, and the standard deviation 17.4.¹

The comparison is what carries the story. Riley and colleagues audited the scoreboard's first five years in 2018 and found no evidence of improvement between 2013 and 2018.² Extending the record by five and a half years and more than tripling the sample moved the mean absolute error from 14.0 hours to 12.8 — a change the reviewers describe as small and of uncertain statistical significance.¹ A separate verification of 273 real-time runs of the WSA–Enlil model between 2010 and 2016 had put the error at 10.4 hours.⁴

Exhibit 01

Every established model lands in the same band

Source: Kay et al. (2024), Table 3 — the eleven model sets with the most scoreboard submissions, together 85.5% of the 1,702 matched predictions. Bars are mean absolute error in hours over the full 2013–23 window, not a like-for-like event set: each group chooses which eruptions to submit for.¹

Key finding. The best-performing large set, an ensemble run by NASA's Moon to Mars office, averages 10.3 hours; the worst, 14.4. The gap between the best and worst well-established forecasting operations is smaller than the error either of them carries.

The upgrade that did not land

Better inputs, worse hindcasts

In 2017 NOAA's Space Weather Prediction Center and the modelling centre began a controlled trial of the obvious remedy: take 38 historical eruptions, hold the measured CME parameters fixed, and re-run them through successively upgraded versions of the same model — newer code, zero-point-corrected magnetograms, and a data-assimilating model of the solar surface. The final report, released in December 2025, covers 1,292 simulations against 42 observed arrivals.³

Over the full event set the upgrades did not help. The replicated pre-2019 operational configuration scored a mean absolute error of 6.6 hours; the configuration pairing the newest model versions with corrected magnetograms scored 9.0, or 2.4 hours worse. Only on a five-event subset from 2017–19, where the magnetogram corrections were reliable, did the upgraded chain beat the benchmark — and the report states plainly that five events cannot settle it.³

What has measurably changed is the observing rather than the modelling. NOAA's SOLAR-1 spacecraft was declared fully operational in June 2026 and has replaced DSCOVR in the solar-wind feed.⁵ But it sits roughly 1.5 million kilometres upstream, which converts a day-scale forecast into a high-confidence warning issued minutes to a couple of hours ahead.⁶ The scoreboard reviewers point instead to consistent stereoscopic viewing, which they expect from the early-to-mid 2030s.¹

What we don't know

The scoreboard is voluntary, so it is not a complete forecast record: groups choose which eruptions to model, and about a third of all submissions were for CMEs never observed to arrive at all — a false-alarm rate the data cannot properly test, because a non-submission is not a prediction of no impact.¹ Errors are scored against the observed shock, while some models predict the slower ejecta behind it, which biases those entries late by an amount nobody has separated out.¹ The NOAA trial re-ran hand-picked events as hindcasts with fixed inputs, so it isolates the model and its magnetograms and says nothing about a forecaster's real-time judgement.³ And neither record establishes why the error is stuck: the reviewers' best inference is that uncertainty in reconstructing the eruption itself is large enough to mask everything downstream of it, but that is an inference, not a measurement.¹

By the numbers

10.1h

Median absolute error, all 1,702 matched predictions

88.7h

Largest single arrival-time error in the record — 3.7 days

35%

Share of submissions for CMEs never observed to arrive

0.1h

Mean bias of NOAA SWPC's entries, the smallest of the large sets

Caveat. An error measured in hours conflates fast and slow eruptions, and the same review finds error rising with transit time. Normalised by transit time, the record moves more than the headline figure suggests: roughly 20% of transit time in 2013–15 against roughly 15% in 2022–23, on the authors' own reading. Both measures come from the same dataset, and the choice between them is a judgement about what a forecast user needs — a fixed hour budget, or a proportional one.¹

Method. Arrival-time error is the predicted arrival minus the observed arrival; mean absolute error (MAE) is the average of those differences ignoring sign, and is the metric both the scoreboard review and the NOAA trial report. Scoreboard figures cover 15 May 2013 to 31 December 2023 and are restricted to the 1,702 submissions matching an observed arrival; the 915 submissions with no observed arrival are excluded from every error figure and reported separately. Exhibit 01 shows the eleven model sets the review identifies as the most frequently submitted; the 85.5% share and the 35% false-alarm share were recomputed in Node from the published counts and match the source. NOAA trial figures are the all-years column (42 arrivals, 2012–2019) of the report's validation table; they are hindcasts with CME input parameters held fixed, not real-time forecasts. No projected or modelled-scenario figure appears in the headline, standfirst or any hero stat.

Sources. Numbered, highest tier first; each inline citation points here.

T1Kay, C., Palmerio, E., Riley, P., et al. (2024), “Updating Measures of CME Arrival Time Errors”, Space Weather 22, e2024SW003951 — peer-reviewed analysis of the CCMC Arrival Time Scoreboard, 2013–2023. agupubs.onlinelibrary.wiley.com. Retrieved 5 Aug 2026.
T1Riley, P., Mays, M. L., Andries, J., et al. (2018), “Forecasting the Arrival Time of Coronal Mass Ejections: Analysis of the CCMC CME Scoreboard”, Space Weather 16, 1245 — the first five-year audit. arxiv.org/abs/1810.07289. Retrieved 5 Aug 2026.
T1Mays, M. L., MacNeice, P. J., Adamson, E. T., et al. (2025), “NASA/NOAA MOU Annex Final Report: Evaluating Model Advancements for Predicting CME Arrival Time” — joint NOAA SWPC and NASA CCMC project report, version 17 December 2025. arxiv.org/abs/2512.14462. Retrieved 5 Aug 2026.
T1Wold, A. M., Mays, M. L., Taktakishvili, A., et al. (2018), “Verification of real-time WSA-ENLIL+Cone simulations of CME arrival-time at the CCMC from 2010–2016”, J. Space Weather Space Clim. 8, A17. arxiv.org/abs/1801.07818. Retrieved 5 Aug 2026.
T1NOAA Space Weather Prediction Center, “SOLAR-1 CCOR-2 Now Fully Operational and on Our Webpage”, 23 July 2026, and “Solar Wind Data and Display Changes”, 30 June 2026. spaceweather.gov. Retrieved 5 Aug 2026.
T1NOAA Space Weather Prediction Center, “Subscription Services” — product definitions and lead times for geomagnetic storm watches and warnings. swpc.noaa.gov. Retrieved 5 Aug 2026.

Corrections. v1.1 (20 Oct 2026): on phone screens the dashed reference line in Exhibit 1 ran through the chart’s value labels; the labels now sit on top of it. No figures changed.

Figures to 5 Aug 2026. Noodle News · v1.1 · cme-arrival-forecast-error.