How Well Does Our Radar Hail Record Match What People Saw on the Ground?

What we compared, and why it's a fair test

On one side are radar records: what NOAA's weather radars estimated about hail inside a storm. On the other are people: weather spotters, emergency managers, members of the public and volunteer observers who reported hail where they were. We asked a simple question of each report: did the radar layer show hail near that spot on that date?

The radar side is the radar layer HailScore™ relies on, taken exactly as NOAA publishes it. This page does not grade HailScore's score or the sizes shown on a HailScore report. When HailScore builds a report, it can replace a radar size estimate with a nearby ground report's measured size. Grading those blended sizes against the same reports that fed them would be circular, so they are not used here on either side.

Each table in our database comes from one public publisher, loaded as published: NEXRAD hail signatures from NOAA's Severe Weather Data Inventory, MRMS estimates from NOAA's MRMS grid, Storm Events from NOAA's National Centers for Environmental Information, volunteer reports from CoCoRaHS, and preliminary reports from the Storm Prediction Center. No radar record is made from a ground report, and no ground report is made from radar.

The ground sets are kept apart and never added together. They overlap: 200 Storm Events reports in this window list CoCoRaHS as their source. The SPC Colorado hail archive (1955–2023) is left out, because it largely repeats Storm Events.

Dates are Mountain Time calendar dates, because a Colorado evening storm falls on the next day in UTC. A radar record counts when it is 0.75″ or larger: that is HailScore's own floor, the one its date check uses, and it is not a National Weather Service threshold. The NWS severe-hail criterion is 1.00″, which we use for the misses list and for radar-only days. Distances are great-circle miles from the reported point to each radar record.

Results: how often radar showed hail near a ground report

A report counts as matched when there is a NEXRAD radar hail record of 0.75″ or larger within 5 miles on the same Mountain Time date. Reports from months when our NEXRAD archive was incomplete are left out of these tables and listed further down; a gap in our copy of the radar record is not a radar miss.

MRMS, in its own window

MRMS is a separate radar layer: NOAA's national grid of hail-size estimates. Our MRMS archive starts in May 2026, so these figures cover May 11, 2026 to September 25, 2026 and nothing earlier. They are not yearly figures, and every one is a small window.

How far away was the nearest radar record?

NEXRAD records a hail signature at the center of a storm cell, once per radar scan, not at the spot where stones landed. A hailstorm's footprint can run for miles from that center, so a report near the edge of a swath can sit a few miles from the nearest record even when the radar saw that very storm.

That is why the 1-mile rule matches far fewer reports than the 5-mile rule. The table shows, for every scored report, how far away the nearest NEXRAD record of 0.75″ or larger was on the same date.

How close were the size estimates?

For each matched report we took the largest radar estimate within the rule's distance on that date and subtracted the reported stone. A positive number means radar estimated larger hail than the person reported. Taking the largest estimate over a 5-mile circle pushes the difference up by construction, so we also show the same figures within 1 mile, which we fixed in advance as the secondary view.

NEXRAD reports sizes in quarter-inch steps and shows nothing above 4.00″: NOAA's product specification for the WSR-88D (ICD 2620003AE, §19.2.2.1) displays any larger estimate as more than 4.00″. A 4.00″ record therefore means "4.00″ or more". We decided before seeing any size result to leave those pairs out of the median and the middle half, and to count them on their own.

Every miss

A miss is a ground report of 1.00″ or larger (the NWS severe-hail criterion) with no radar record of 0.75″ or larger within 5 miles on the same date. Each row links to its public record. "Nearest radar that date" is the closest qualifying record anywhere in the area we read, which is Colorado plus about ten miles past the state line.

Reports not scored against NEXRAD

These reports are not scored against NEXRAD at all, because our copy of the NEXRAD record held less than 95% of what NOAA publishes for Colorado that month (see Limits). They are neither matches nor misses.

Why radar can miss a ground report

• The radar beam rises with distance and can pass over the part of a storm where hail forms, or be blocked by mountains.
• Small or short-lived cells can grow and drop hail between scans, or be too small for the algorithm to flag.
• Timing: a storm just before or after midnight can land on different dates in the two records.
• Location: a report's coordinates can be approximate, sometimes to the nearest town.
• These are general reasons radar can miss hail. We do not diagnose individual rows unless the data shows the cause, as the notes column does when no NEXRAD record exists anywhere that day.

Radar-only days (no ground report)

Here the question runs the other way: on how many county-days did NEXRAD show hail of 1.00″ or larger with no report of any size within 5 miles that date? We call these radar-only (no ground report). They are not errors by default.

Ground reports are sparse. They come from where people live and are watching, so storms over open plains, at night or in rural counties often go unreported. Each ground set is counted on its own.

Counties are placed with Census county boundaries. A few months when our NEXRAD archive was incomplete are included here; missing radar records can lower these counts, never raise them.

Stop rules that tripped, and what we checked

Before counting anything, we wrote down results that would make us stop and look for a mistake. On the first run, 8 of them tripped:

1. NEXRAD vs NOAA Storm Events reports, 2022: 191 of 198 reports matched. That is outside the band we set in advance (between 5% and 95% for any year with at least 30 reports).
2. NEXRAD vs NOAA Storm Events reports, 2023: 740 of 769 reports matched. That is outside the band we set in advance (between 5% and 95% for any year with at least 30 reports).
3. NEXRAD vs NOAA Storm Events reports, 2026: 240 of 245 reports matched. That is outside the band we set in advance (between 5% and 95% for any year with at least 30 reports).
4. NEXRAD vs NOAA Storm Events reports: within 1 mile, same date, 1,942 of 4,571 matched, against 4,184 of 4,571 under the main rule. A sensitivity run moving the result by more than half was a stop we set in advance.
5. NEXRAD vs CoCoRaHS volunteer reports: within 1 mile, same date, 285 of 722 matched, against 613 of 722 under the main rule. A sensitivity run moving the result by more than half was a stop we set in advance.
6. NEXRAD vs SPC preliminary reports, 2026: 51 of 51 reports matched. That is outside the band we set in advance (between 5% and 95% for any year with at least 30 reports).
7. NEXRAD vs SPC preliminary reports: no report of 1.00″ or larger went unmatched (51 such reports scored). We set an empty misses list as a stop in advance.
8. MRMS vs SPC preliminary reports, 2026: 82 of 84 reports matched. That is outside the band we set in advance (between 5% and 95% for any year with at least 30 reports).

What we checked

• Placebo test, added after the first run (see Method changes below). We ran the main rule again with every NEXRAD date moved later, so each report was checked against radar from a different day. If radar showed hail near most places on most days, the moved dates would still match. They mostly did not:
  • NOAA Storm Events reports, NEXRAD dates moved 7 days later: 262 of 4,493 reports matched (95% interval 5.2% to 6.6%; 78 reports left out because the moved date fell outside our NEXRAD window or in an incomplete month).
  • NOAA Storm Events reports, NEXRAD dates moved 365 days later: 164 of 3,999 reports matched (95% interval 3.5% to 4.8%; 572 reports left out because the moved date fell outside our NEXRAD window or in an incomplete month).
  • CoCoRaHS volunteer reports, NEXRAD dates moved 7 days later: 35 of 705 reports matched (95% interval 3.6% to 6.8%; 17 reports left out because the moved date fell outside our NEXRAD window or in an incomplete month).
  • CoCoRaHS volunteer reports, NEXRAD dates moved 365 days later: 18 of 640 reports matched (95% interval 1.8% to 4.4%; 82 reports left out because the moved date fell outside our NEXRAD window or in an incomplete month).
• The one-mile drop. NEXRAD records storm-cell centers, so the one-mile rule fails for reports that sit a few miles from the center of the storm that hit them. The distance table above shows where the reports fall.
• The SPC preliminary reports are a small, recent sample (51 reports against NEXRAD, 84 against MRMS), all of 1.00″ or larger, from late-summer storms.
• Every Storm Events date was computed twice, in the database and in the comparison code, and the script stops on any disagreement. There were none. The rules that pair a report with a radar record are covered by automated tests.

We found no error in the matching, so the results are published as they came out, with the method unchanged.

What this page does not measure

• Roof damage. A radar record or a ground report of hail near a home says nothing about whether a roof was harmed. HailScore has not published any validation against inspected roofs: the count of usable outcome records is zero, as our methodology's validation appendix states.
• Insurance outcomes. Nothing here is about claims, approvals or payments.
• Other states. Every comparison on this page covers Colorado and no other state.
• HailScore's score or report sizes. This page tests the radar layer underneath them, not HailScore's own output.

The full method behind HailScore, including its validation status, is in the methodology white paper. What radar can and cannot establish is covered in what radar can and cannot prove.

Limits

• Our copy of the NEXRAD record is not complete. We compared it with what NOAA serves for the same Colorado area, one month at a time, from January 2015 to August 2026: NOAA 407,802 records, ours 396,810. Months where we held less than 95% are listed below and left out of every NEXRAD rate.
• The MRMS window is short: May 11, 2026 to September 25, 2026. MRMS figures are small samples and are never yearly figures.
• NOAA publishes Storm Events months after a storm. Our Storm Events window ends on June 30, 2026, the latest month NOAA had released when we ran this.
• Ground reports follow people. Cities, highways and daytime storms are reported far more often than open country or night storms, so every match rate here describes reported hail, not all hail.
• Radar size estimates lose skill for very large hail, roughly above three inches, as our methodology notes; NEXRAD also stops at 4.00″ (see the size section).
• These figures are a snapshot as of 2026-10-06 11:34 MDT. Late reports can still arrive for dates inside these windows, so a later run can differ slightly.

Download and reproduce

Colorado ground hail reports with no radar match (misses) (CSV)

Every ground report of 1.00″ or larger that had no radar record under the main rule, for each radar layer and ground set on this page, with its public record link. No coordinates are included; follow each record link for the location. Free to download and reuse under CC BY 4.0 with attribution to HailScore; CoCoRaHS asks that users acknowledge it as a source.

License: CC BY 4.0. Columns: radar_layer, ground_source, record_id, record_url, mt_date, county, nearest_place, ground_size_in, report_source, nearest_radar_detection_mi, radar_below_floor_within_5mi, radar_rows_that_date.

Rules used

• Detection floor: 0.75″ (HailScore's own, not an NWS threshold)
• Severe threshold for misses and radar-only days: 1.00″ (NWS)
• Main rule: within 5 miles, same date; alternative rules: within 1 mile, same date; within 5 miles, same date or the day before or after; within 10 miles, same date
• Dates: America/Denver calendar dates
• Size classes: 0.75″ to under 1.00″, 1.00″ to under 1.75″, 1.75″ or larger
• Intervals: 95% Wilson score; cells with fewer than 30 reports are marked small sample
• NEXRAD archive months below 95% of NOAA's count are not scored
• MRMS: each date's largest value per grid cell, minus cells that merely repeat the previous day's 24-hour maximum (113,176 such repeats dropped)

Data windows and cutoffs

• NOAA Storm Events: January 1, 2015 to June 30, 2026
• CoCoRaHS: January 1, 2015 to September 25, 2026
• SPC preliminary reports: July 21, 2026 to September 22, 2026
• NEXRAD: January 1, 2015 to August 31, 2026
• MRMS: May 11, 2026 to September 25, 2026

Generated 2026-10-06 11:34 MDT. Anyone can repeat this with the public NOAA, CoCoRaHS and SPC records listed under Sources and the rules above.

Update log

• October 6, 2026: First published. Every figure comes from one run of the comparison script, and every rule was written down before the first match was counted.

Decisions made before any result was seen

• October 6, 2026: NEXRAD's 4.00-inch value is treated as '4.00 inches or more' (censored): left out of the median and IQR of the size difference, counted on its own, and counted in 'radar at or above ground' only when the ground stone was 4.00 inches or smaller. Why: The WSR-88D Hail Index product shows any size above 4.00 inches as '>4.00' (ICD 2620003AE §19.2.2.1), and SWDI publishes no value above it.

Method changes after first run

• October 6, 2026: Added a placebo check: the primary rule (5 miles, same date, 0.75-inch floor) run again against NEXRAD records from 7 days and 365 days after each report's date. It changes no published rate. Why: The first run tripped pre-registered stop rules, including source-years above 95%. A rule that matched because radar shows hail somewhere nearby on most days would still match after the shift.
  • NOAA Storm Events reports, NEXRAD dates moved 7 days later: 262 of 4,493 reports matched (95% interval 5.2% to 6.6%; 78 reports left out because the moved date fell outside our NEXRAD window or in an incomplete month).
  • NOAA Storm Events reports, NEXRAD dates moved 365 days later: 164 of 3,999 reports matched (95% interval 3.5% to 4.8%; 572 reports left out because the moved date fell outside our NEXRAD window or in an incomplete month).
  • CoCoRaHS volunteer reports, NEXRAD dates moved 7 days later: 35 of 705 reports matched (95% interval 3.6% to 6.8%; 17 reports left out because the moved date fell outside our NEXRAD window or in an incomplete month).
  • CoCoRaHS volunteer reports, NEXRAD dates moved 365 days later: 18 of 640 reports matched (95% interval 1.8% to 4.4%; 82 reports left out because the moved date fell outside our NEXRAD window or in an incomplete month).