Wind and Hail Risk from Public Storm Records

Every commercial property in the United States sits inside a free, seventy-year national record of the wind and hail that has hit near it. The record is published by NOAA, it costs nothing, it downloads in minutes, and it is the layer most lenders reach for when a borrower asks whether a roof in Oklahoma or a tilt-up in Alabama has a storm problem. It is also, read carelessly, one of the most misleading datasets in the public stack.

The problem is not that the data is bad. NOAA is unusually candid about what it is and what it is not. The problem is that the record answers a narrower question than the one a credit file asks. It tells you what someone reported, where they reported it, and how confident the reporter was. It does not tell you the wind load a building was designed for, the frequency of a peril at a point, or what a loss would cost. Those are three different questions with three different owners, and only the first belongs to the storm record.

This article is about reading that record honestly: what the public storm databases actually contain, the conventions and discontinuities inside them, and what a lender may and may not conclude from a point-radius storm history. It sits inside the public-data stack for commercial real estate analysis as the layer with the most generous coverage and the weakest measurement basis. The national screening method built on top of it, the one that turns these files into a map rather than a property answer, is set out separately in the national wind risk screening methodology.

One boundary governs everything that follows. Public storm records are observations, not a loss model. Reading storm climatology from them supports screening and context. It never replaces engineering wind design, which is governed by the building code through ASCE 7, and it never replaces insurance pricing. The decisions rest with the lender, its engineers and its insurers.

The five public records, and what each one is

"The storm record" is not one file. It is five distinct products with different producers, different windows, different grain and different purposes, and a screen that does not say which one it used is not reproducible.

The NOAA NCEI Storm Events Database is the archive of record. NCEI describes it as containing "the records used to create the official NOAA Storm Data publication", covering January 1950 to May 2026, as entered by the National Weather Service. Storm Data itself documents the occurrence of storms and other significant weather phenomena "having sufficient intensity to cause loss of life, injuries, significant property damage, and/or disruption to commerce", plus rare phenomena that generate media attention and other significant meteorological events. It is distributed through a web search interface and as annual bulk CSV files, currently seventy-seven annual detail files in the public directory, each stamped with a data year and a creation date (NOAA NCEI, 2026).

The Storm Prediction Center severe weather database is the working extraction of the same underlying reports for the three convective perils. It publishes tornado records from 1950 to 2025, hail from 1955 to 2025 and damaging wind from 1955 to 2025, in comma-separated files by year, by period and as complete-record archives. Its own format specification is explicit about the relationship: these files "are an attempt to represent the data that is submitted to the Storm Data publication by National Weather Service field offices" (NOAA SPC, 2010). For anyone doing arithmetic rather than looking up one address, this is the more tractable of the two.

National Weather Service local storm reports are the raw feed. They are the real-time messages that a forecast office issues as reports arrive, and they are what eventually becomes a Storm Data entry after review. They matter to a lender mainly as an explanation of provenance: the archive is downstream of a real-time operational product, not of a measurement programme.

HURDAT2 is the tropical cyclone best-track database from the National Hurricane Center. The Atlantic file covers 1851 to 2025 and the northeast and north central Pacific file covers 1949 to 2025, both last updated on 27 February 2026. Each is comma-delimited text carrying six-hourly position, maximum winds, central pressure and, beginning in 2004, size (NOAA NHC, 2026). It is a track database, not a damage database, and it is the only one of the five that has been systematically reanalysed.

ASCE 7 is not a storm record at all. It is the design standard that the building code adopts, and the free ASCE Hazard Tool returns its values for a site: wind, tornado, seismic, ice, rain, flood, snow and tsunami parameters for ASCE/SEI 7-10, 7-16, 7-22, 41-17 and 41-23. For wind it returns three-second gust speeds at 33 feet above ground for Exposure Category C, by risk category (ASCE, 2026). It belongs in this list because it is the reference a wind screen must be measured against, and because confusing it with the observational record is the single most common error in this area.

A sixth product deserves a mention because lenders often meet the storm record through it rather than directly. FEMA's National Risk Index, version 1.20 of December 2025, publishes expected annual loss, social vulnerability and community resilience for eighteen hazards at county and census tract level, and four of those eighteen are hail, strong wind, tornado and hurricane (FEMA, 2025). It is a useful summary layer. It is not an escape from the caveats below, because its convective hazard components are themselves built on the same reports.

The database changed shape three times

The single most important fact about the Storm Events Database is that it is not one series. NCEI sets out the periods plainly. From 1950 to 1954, only tornado events were recorded. From 1955 to 1992, only tornado, thunderstorm wind and hail events were keyed from the paper publications into digital data. From 1993 to 1995, only those same three event types were extracted from the unformatted text files. From 1996 to the present, forty-eight event types are recorded, as defined in National Weather Service Directive 10-1605 (NOAA NCEI, 2026).

NCEI states the consequence in its own words: "Due to changes in the data collection and processing procedures over time, there are unique periods of record available depending on the event type." It also records something a data user needs to know about the archive's own hygiene. NCEI standardised the event types across the record, but it "has not changed any data values for locations, fatalities, injuries, damage, narratives and any other event specific information". The archive was harmonised at the label level and left alone at the value level.

For a lender the arithmetic is unforgiving. A screen that reports "events near this property since 1950" is reporting a count assembled under four different collection regimes, three different capture technologies and two different definitions of what counts as an event. Any long-run trend drawn from it is measuring administrative history at least as much as weather. The defensible window for anything other than tornado, hail and thunderstorm wind is 1996 forward, which as of the end of 2025 is thirty complete years.

Even within the three long-running perils the rules moved. The National Weather Service changed the criterion for severe hail from a 0.75 inch minimum to a 1.00 inch minimum in 2010, and SPC notes that for legacy purposes 0.75 inch reports continue to be included in the annual hail files (NOAA SPC, 2026). In 2016, estimated F-scale ratings were calculated for more than 1,800 tornadoes previously carried in the database with an F-scale entry of minus nine, meaning unknown, and a new field was added so those records can be identified. SPC also notes that tornadoes were likely underreported prior to 1953, and that many tornado statistics are therefore derived from 1953 or later.

None of this is hidden. All of it is on the pages that serve the data. It is simply not carried forward into the screens that get built on top, which is how a count that is really a record of changing definitions arrives in a credit memorandum as a statement about weather.

Most of the wind record is an estimate

Here is the part that changes how a wind screen should be read, taken from inside the database rather than from argument.

Every severe thunderstorm wind report in the national record carries a magnitude in knots and, since the mid-2000s, a code saying how that magnitude was obtained. The codes are EG for an estimated gust, ES for an estimated sustained wind, MG for a measured gust and MS for a measured sustained wind (NOAA NCEI, 2026). The directive that governs data entry is direct about the default: "If known, maximum gusts will be encoded as 'measured'; otherwise, they will be an estimate" (NWSI 10-1605, 2021).

MMCG tabulated the Storm Prediction Center's complete damaging wind file on 25 August 2026. Across 2006 to 2025, the twenty years in which every record carries a magnitude-type code, the file holds 334,594 severe wind reports. Of those, 289,837 record an estimated wind speed and 44,757 record a measured one. Estimates are 86.6 percent of the national severe wind record. Measurements are 13.4 percent.

That alone would be worth knowing. What makes it decisive is where the estimates land. Among the 289,837 estimated reports, 115,146 carry exactly 50 knots and 68,108 carry exactly 52 knots. Fifty knots is the severe threshold itself, the 58 mph at or above which a thunderstorm wind event qualifies for the record. Fifty-two knots is 60 mph. Those two values account for 63.2 percent of every estimated wind report in twenty years.

The measured subset behaves completely differently, and that is the proof. Among the 44,757 measured reports, 50 knots accounts for 12.5 percent and 52 knots for 12.3 percent, together under a quarter of the group. The values immediately adjacent tell the story even more clearly. At 51 knots the estimated series collapses to 0.7 percent while the measured series holds at 10.1 percent. At 53 knots the estimated series is 0.9 percent and the measured series 7.8 percent. Instruments produce a smooth distribution across 50, 51, 52 and 53 knots. Human estimates produce two spikes and two holes.

The gap runs in a consistent direction as well. Measured reports have a median of 54 knots and a mean of 56.0 knots against 52 and 53.5 knots for estimates, and 9.3 percent of measured reports reach 65 knots or more against 6.6 percent of estimates. Where an anemometer exists, the wind is recorded as somewhat stronger and considerably more varied than where a person judged it from what the wind did to a tree.

None of this makes the record useless. A report is still evidence that a severe convective wind event occurred near a place on a date, and that is genuinely valuable. What it means is that the magnitude column is not a measurement series and cannot be treated as one. Averaging it, ranking sites by it, or reporting a "maximum recorded gust" for a property is arithmetic performed on a field that is, five times out of six, a judgement made at the roadside. The honest output of a wind screen is a count of reported events with a window and a radius attached, not a wind speed.

The measured share is also rising, which is its own homogeneity problem. It ran 8.0 percent in 2006 and 21.2 percent in 2025. The basis on which the national wind record is compiled has been changing continuously underneath anyone comparing decades.

Hail size is a category dressed as a measurement

Hail is recorded in inches to the hundredth, which reads like precision and is not. The Storm Data directive states the convention and, in doing so, gives the game away: "Hail size will be given in hundredths of an inch (0.50, 0.75, 0.88, 1.00, 1.50, etc., are the most common)" (NWSI 10-1605, 2021). Those are not arbitrary hundredths. They are the diameters of familiar reference objects, and a spotter reports the object rather than a measurement.

The complete SPC hail file holds 414,481 reports from 1955 to 2025. MMCG tabulated the size column on 25 August 2026. Five values carry 87.9 percent of the entire record: 1.00 inch at 122,540 reports, 0.75 inch at 104,310, 1.75 inch at 68,664, 0.88 inch at 49,300 and 1.50 inch at 19,387. A field documented to the hundredth of an inch is, in practice, a five-level ordinal variable.

This matters for commercial property because hail underwriting turns on size. A roof system's vulnerability is not linear in hailstone diameter, and the difference between a 1.00 inch and a 1.75 inch event is the difference between cosmetic granule loss and a claim. A screen that reports the maximum hail size within a radius is reporting the largest reference object anybody happened to compare a stone to, from whichever report happened to be filed. It is a lower bound on what fell, recorded with a precision the observation method cannot support.

The 2010 threshold change compounds it. Because the severe criterion moved from 0.75 inch to 1.00 inch, the share of hail reports below one inch fell from 53.5 percent in the 2000 to 2004 period and 51.1 percent in 2005 to 2009 to 28.5 percent in 2010 to 2014, 22.8 percent in 2015 to 2019 and 17.3 percent in 2020 to 2024. Nothing about hailstorms changed on that date. The definition of a reportable event did. A property whose hail history is counted from 1996 is being counted under two different rules, and the earlier half of the window will look busier for reasons that have nothing to do with the site.

Counting reports is not counting storms

The unit inside these files is a report, and a report requires a person. That single dependency generates most of the remaining distortions.

Start with who reports. NCEI states that the National Weather Service receives its information from "county, state and federal emergency management officials, local law enforcement officials, skywarn spotters, NWS damage surveys, newspaper clipping services, the insurance industry and the general public, among others". The SKYWARN spotter network is a volunteer programme of between 350,000 and 400,000 trained observers whose training is free and runs about two hours (NOAA NWS, 2026). This is a serious and effective public-service network. It is not an instrument grid, and it is distributed the way people are distributed.

The consequence is the reporting-density bias, and it is not folklore: it is visible in the record's own structure. Reports arise where there are people to see them, roads to see them from and property to be damaged. A hailstorm crossing open rangeland and the same storm crossing a county seat produce different numbers of records, and the difference is a population map, not a weather map. At fine geographic grain, a raw report-density surface tracks settlement as much as meteorology, which is why the national methodology piece normalises before it maps.

The time dimension shows the same thing cleanly. MMCG tabulated the SPC tornado file, 73,458 records from 1950 to 2025, by damage rating and five-year period on 25 August 2026. Average annual F0 tornado reports, the weakest category and the one most dependent on somebody noticing, rose from 50 a year in 1950 to 1954 to 853 a year in 2000 to 2004, a seventeen-fold increase. Over the same span, tornadoes rated F2 and above, which leave damage a survey team can find days later, went from 153 a year to 113. The two series behave nothing alike.

Honesty requires the reverse caveat too. The F2-and-above series is not a clean control. Rating practice changed as well, the Enhanced Fujita Scale took effect on 1 February 2007 (NOAA SPC, 2026), and the F0 series has itself fallen since the mid-2000s peak. The correct conclusion is not that the strong-tornado series is truth and the weak one is noise. It is that neither series is homogeneous, and that a chart of total reports over time is a chart of observing effort with a weather signal inside it.

Three structural rules complete the picture. First, the directive tells preparers that where a continuous or nearly continuous swath of thunderstorm wind or hail damage occurred, "a single event should be entered into Storm Data", because a swath "reduces the chance of a researcher interpreting a single event as a series of events occurring across multiple points". Whether one storm becomes one record or six is therefore a preparer's judgement. Second, the database is organised as episodes containing events: an episode is an entire storm system, an event is an individual storm type within it, and events in the same episode may begin no more than five calendar days apart. Third, in the SPC tornado files a tornado crossing state lines or more than four counties is written as several records sharing one identifier, and the format specification warns that understanding those fields "is critical to counting state tornadoes, totaling state fatalities/losses". Naive row counts double-count.

Where the event happened, and how precisely

A point-radius screen assumes the record knows where the event was. It half does.

Every Storm Events record carries a geography type: C for a county or parish, Z for an NWS public forecast zone and M for marine (NOAA NCEI, 2026). Those are not the same object. County-based event types, including hail, thunderstorm wind, tornado, flood and lightning, resolve to a county polygon. Zone-based types, including high wind, hurricane, winter storm and excessive heat, resolve to a forecast zone, which may be much larger than a county or may be a subdivision of one. A screen that treats every record as a county record will misplace an entire class of wind events.

Within a county record, the location is expressed the way a person on a telephone describes a location. The fields are a range in miles to the nearest tenth, a sixteen-point compass azimuth and the name of a city, town or village from which the range is calculated. There is also a begin latitude and longitude and an end latitude and longitude for the event or damage path. The point in the file is therefore a derived coordinate for a described position, and its accuracy is the accuracy of "about four miles east-northeast of a named town". For a one-mile radius drawn around a parcel, that is a material uncertainty, and it compounds with whatever error the parcel's own coordinate carries. The same discipline applies here as in geocoding accuracy in CRE analysis: a spatial join is only as good as the weaker of its two coordinates.

The source field offers no rescue. NCEI documents it with examples including Public, Newspaper, Law Enforcement, Broadcast Media, ASOS, Park and Forest Service, Trained Spotter and CoCoRaHS, and then states that it "can be any entry; isn't restricted in what's allowed". It is free text. A screen cannot filter to instrument-sourced reports on this field with any confidence, which is exactly why the magnitude-type code discussed above is the more reliable quality flag for wind.

One further field is worth knowing about because it looks useful and is not. The bulk format documents a "category" field with the note that "During the time of downloading this particular file, NCEI has never seen anything provided within this field." An empty column with a promising name is a standing invitation to a mistake.

The damage columns, and why underwriting cannot use them

Both databases carry property damage and crop damage. Both are unusable as loss evidence, and both producers say so.

The Storm Prediction Center is blunt: "Monetary loss information is highly suspect and should be used with caution, if at all" (NOAA SPC, 2026). Its format specification explains why in structural terms. Prior to 1996 the tornado loss field is not a dollar amount at all but a categorisation, with code 1 meaning under $50, code 5 meaning $50,000 to $500,000, code 9 meaning up to $5 billion and so on. From 1996 the same column holds estimated property damage in millions of dollars. A single column changes units mid-record. The specification adds a warning that closes off the obvious workaround: "Entry of 0 does not mean $0." Zero means unknown.

The Storm Data directive explains how the figures are produced. Property damage estimates should be entered as actual dollar amounts "if a reasonably accurate estimate from an insurance company or other qualified individual is available". If it is not, the preparer has two choices: check the "no information available" box, or make an estimate. Estimates may be obtained from emergency managers, the U.S. Geological Survey, the U.S. Army Corps of Engineers, utility companies and newspaper articles. They are to be rounded to three significant digits.

Two further exclusions matter for anyone tempted to read these fields as a cost of loss. Under the directive, the costs of snow removal, debris clearing and moving, firefighting, personnel overtime and public housing assistance "will not be tallied as directly-related parts of the property/crop damage". And NCEI states that "Damage amounts are not adjusted for inflation and are the values that were entered at the time of the event". A 1998 figure and a 2024 figure sit in the same column in different money.

Put together: a field that is sometimes a category and sometimes a dollar figure, sometimes sourced from an insurer and sometimes from a newspaper, rounded to three digits, excluding much of what an owner actually spends, never inflation-adjusted, and with zero standing for unknown. There is no defensible way to build a loss expectation for a single commercial property from that column. Loss estimation belongs to catastrophe modelling and to insurers, and the public record's contribution to it is the event history, not the dollar figures. This is the boundary the whole article turns on: the storm record documents that something happened, not what it cost.

What the building code asks, and why it is a different question

The clearest way to see the limits of the observational record is to put it beside the standard a building is actually designed to.

ASCE 7 sets design wind speeds through maps keyed to risk category, and each map is drawn to a mean recurrence interval. In ASCE 7-16 those intervals are 300 years for Risk Category I, 700 years for Risk Category II, 1,700 years for Risk Category III and 3,000 years for Risk Category IV, corresponding to target reliability indices of 2.5, 3.0, 3.25 and 3.5 respectively (McAllister, Wang and Ellingwood, 2018). An ordinary commercial building is Risk Category II. Its wind design is calibrated against a 700-year event.

Now set the public record against that. For thunderstorm wind and hail, the consistent record is seventy-one years long and the internally homogeneous portion is shorter still. For all forty-eight event types it is thirty years. At the grain of a single parcel, with a one-mile or three-mile radius, thirty years of reported events is a very thin sample of a rare-event distribution, and the events it does contain carry estimated magnitudes clustered on two values.

These two things cannot be reconciled and should not be. The design standard asks a 700-year question and answers it with a hazard model built from long instrumental records, hurricane simulation and structural reliability analysis. The public storm record answers a thirty-year observational question about reported occurrences near a place. Both are legitimate. Neither substitutes for the other, and a screen that reports "no severe wind events within three miles in twenty years" has said nothing whatsoever about whether the building meets its design wind speed.

The practical instruction follows. Where wind or hail exposure is material to a credit, the engineering question is answered by looking up the site in the ASCE Hazard Tool and, where the exposure justifies it, by asking what the building was actually designed and built to. The observational record is context around that answer, not a replacement for it.

A defensible wind and hail screen for one property

With the caveats established, a workable screen is short and its outputs are modest.

Fix the window and say why. For hail, thunderstorm wind and tornado, 1996 forward is the defensible default because it aligns with the current Storm Data regime, and 2010 forward is the defensible default if hail size distribution is the question, because that is when the severe hail criterion settled. State the window on the face of the output.

Fix the radius and say what it is doing. A one-mile radius around a parcel is answering "has severe weather been reported essentially at this site", and it will return very few records. A ten-mile radius is answering "is this a place where severe convective weather occurs", and it will return many. Neither is wrong. Reporting one without saying which was used is.

Count events, not magnitudes. The output is a count of reported events by peril, by year, within the stated radius and window, with the geography type noted. Do not average the magnitude column. Do not report a maximum gust as though it were measured, and if the maximum is reported, carry its magnitude-type code beside it so a reader can see whether it was estimated.

Normalise before comparing places. Two sites with different report counts may differ in weather or in the number of people nearby who would have called it in. A comparison across markets needs either a normalisation or an explicit acknowledgement that none was applied, and the national approach to that problem is the subject of the national wind risk screening methodology.

Say what the sample supports. Thirty years of records inside a three-mile circle is a small sample of a rare process. It supports statements of the form "severe hail of one inch or more was reported within three miles in eleven of the last thirty years". It does not support an annual probability, a return period, or a ranking of one site above another on a difference of two or three events. Rare-event statistics at a point are thin, and the honest screen says so rather than manufacturing a decimal.

Cross-check against the design and insurance answers. The ASCE 7 design wind speed for the site, the risk category the building falls in, and the terms of the property policy, particularly any separate wind or hail deductible and any cosmetic damage exclusion, are what actually govern the exposure. The storm record tells the lender where to look harder. That is the same role the flood layer plays in reading flood zone designations, and the same discipline the wider pass applies in what to screen before ordering a Phase I.

The portfolio view, where these records are strongest

Everything above is a warning about single-property inference. The same data behaves much better one level up, and this is where a lender gets real value from it.

At portfolio scale the reporting-density bias partly cancels, because a portfolio of commercial buildings is itself distributed where people and property are. The question changes from "what is the hazard at this point" to "how much of my book sits in places where severe convective wind and hail are reported frequently", and that is a question thirty years of county-level counts can answer usefully.

Three portfolio uses are defensible on this data. The first is concentration measurement: the share of exposure, by unpaid principal balance rather than by loan count, in counties in the upper decile of reported severe wind or hail events per year. That is a concentration statement of the same family as the geographic and property-type work in CRE concentration monitoring for community banks. The second is post-event triage: when a significant convective event occurs, the storm record and the local storm reports give a same-week list of which collateral sat inside the reported swath, which is a servicing and outreach tool rather than a loss estimate. The third is the annual review, where a hazard layer refreshed each year is one of the cheapest ways to notice that a property's context has changed, as set out in annual reviews and portfolio surveillance.

What these uses have in common is that they aggregate. None of them asks a single record to carry weight it cannot bear, and all of them survive the estimated-magnitude problem because they count events rather than measuring them.

Putting it in the file

A storm workup that survives a second reading records six things, and each one closes a gap that the raw data leaves open.

Record the database and the extract date. The Storm Events Database is updated continuously and NCEI receives each month's data roughly seventy-five days after the month ends, so a screen run today and the same screen run in three months will return different answers for the same recent period. Record which file or service was queried and when.

Record the window and the reason for it, the radius, and the geography type of the records returned. Record the counts by peril and by year rather than a single total. Record the magnitude-type distribution for any wind figure quoted, because that is the only field that distinguishes a measurement from a judgement. Record explicitly that the damage columns were not used, and why. And record the design-side answer separately: the ASCE 7 edition, the risk category and the design wind speed for the site, kept in its own sentence so that nobody downstream can read an observational count as a structural conclusion.

Two habits separate a good workup from a misleading one. Date every element, which is the general standard set out in the provenance standard and applies with unusual force here because three of the five products in this stack have changed definitions within the analytical window. And keep the observational sentence and the engineering sentence apart in the text. "Fourteen severe hail events were reported within three miles between 1996 and 2025" and "the design wind speed for this site under ASCE 7-16, Risk Category II, is the figure returned by the ASCE Hazard Tool on this date" are both useful. Collapsing them into a single paragraph is how a report count becomes an implied structural opinion somewhere between the analyst and the committee.

It is also worth stating plainly what this layer cannot be asked to do. It does not know the roof age, the fastener schedule, the deductible structure or whether the last hail claim was paid. It does not know what fell on a parcel where nobody was standing. Its magnitude column is mostly a set of careful guesses, and it says so. What it does know, and knows better than any paid alternative, is where in the United States people have been reporting severe convective weather for seventy years, and that is a genuinely valuable thing for a lender to be able to see for free. The broader public-records pass this screen belongs to is set out in the due diligence data stack.

MMCG Analytics is a map-first commercial real estate analytics platform for lenders and investors, built by MMCG Invest, LLC of San Francisco on federal, state and public data with source and vintage provenance carried on displayed values. Wind risk is one of its analytical layers, sourced from public records; decisions rest with the lender, its engineers and its insurers.

Frequently asked questions

What is the NOAA Storm Events Database?

It is the national archive of the records used to create the official NOAA Storm Data publication, covering January 1950 to May 2026 as entered by the National Weather Service. It documents storms and other significant weather phenomena of sufficient intensity to cause loss of life, injuries, significant property damage or disruption to commerce. It is published through a web search interface and as annual bulk CSV files, and the Storm Prediction Center publishes a parallel extraction of the tornado, hail and damaging wind records in simpler files.

How far back does reliable wind and hail data go?

It depends on the peril. Only tornado events were recorded from 1950 to 1954. Tornado, thunderstorm wind and hail were the only types captured from 1955 through 1995. The current set of forty-eight event types dates from 1996, under National Weather Service Directive 10-1605. For anything other than those three convective perils, 1996 is the earliest defensible start, which is thirty complete years to the end of 2025. Even within the three long-running perils, the severe hail criterion changed from 0.75 inch to 1.00 inch in 2010.

Are the wind speeds in the storm database measured?

Mostly not. Records carry a magnitude-type code distinguishing an estimated gust from a measured one, and the Storm Data directive states that gusts are encoded as measured only if known and are otherwise an estimate. Across 2006 to 2025 the Storm Prediction Center's damaging wind file holds 334,594 reports, of which 289,837, or 86.6 percent, are estimated and 44,757, or 13.4 percent, are measured. Among the estimates, 63.2 percent carry one of two values: 50 knots, which is the severe threshold, and 52 knots, which is 60 mph.

Can I use the damage figures in the storm database for underwriting?

No. The Storm Prediction Center states that monetary loss information is highly suspect and should be used with caution, if at all. In its tornado file the loss column is a damage category code before 1996 and a figure in millions of dollars from 1996, and a zero entry means unknown rather than $0. NCEI states that damage amounts are not adjusted for inflation. The Storm Data directive allows estimates from sources including newspaper articles, requires rounding to three significant digits, and excludes debris clearing, firefighting and overtime costs from the property damage figure.

Does the storm record tell me the wind a building was designed for?

No, and the two should never be mixed. Design wind speeds come from ASCE 7 through the building code, and the free ASCE Hazard Tool returns them as three-second gust speeds at 33 feet above ground for Exposure Category C by risk category. ASCE 7-16 draws those maps to mean recurrence intervals of 300 years for Risk Category I, 700 for Risk Category II, 1,700 for Risk Category III and 3,000 for Risk Category IV. The observational record covers thirty consistent years of reported events. It is context, not a design or insurance input.

Why do storm reports cluster near towns and roads?

Because a report requires an observer. NCEI receives reports from emergency management officials, law enforcement, SKYWARN spotters, NWS damage surveys, newspaper clipping services, the insurance industry and the general public, and the SKYWARN network alone runs to between 350,000 and 400,000 trained volunteers. Reports therefore arise where people, roads and property are. The clearest evidence is in the tornado record: average annual reports of the weakest F0 tornadoes rose from 50 a year in 1950 to 1954 to 853 a year in 2000 to 2004, while tornadoes rated F2 and above went from 153 a year to 113 over the same period.

What radius and window should a property-level storm screen use?

State both explicitly, because they determine the answer. A one-mile radius asks whether severe weather has been reported essentially at the site and returns few records; a ten-mile radius asks whether the area experiences severe convective weather and returns many. For the window, 1996 forward aligns with the current event-type regime and 2010 forward is the cleaner choice for hail size questions. The output should be a count of reported events by peril and by year, with the geography type noted, and not an average magnitude or an implied annual probability.

Sources

1. NOAA National Centers for Environmental Information, Storm Events Database, landing page, cited for the database holding the records used to create the official NOAA Storm Data publication, the January 1950 to May 2026 period, what Storm Data documents, and the statement that NCEI standardised event types without changing data values for locations, fatalities, injuries, damage or narratives; read 25 August 2026. https://www.ncei.noaa.gov/stormevents/
2. NOAA National Centers for Environmental Information, Storm Events Database, database details and scope page, cited for the four collection regimes (tornado only 1950 to 1954; tornado, thunderstorm wind and hail keyed from paper publications 1955 to 1992; the same three extracted from unformatted text files 1993 to 1995; 48 event types from 1996 under NWS Directive 10-1605) and the unique-periods-of-record statement; read 25 August 2026. https://www.ncei.noaa.gov/stormevents/details.jsp
3. NOAA National Centers for Environmental Information, Storm Events Database FAQ, cited for the list of reporting sources, the statement that the NWS does not guarantee the accuracy or validity of the information, the statement that damage amounts are not adjusted for inflation, the approximately 75 day receipt lag and the episode versus event definitions; read 25 August 2026. https://www.ncei.noaa.gov/stormevents/faq.jsp
4. NOAA National Centers for Environmental Information, Storm Data Bulk Data Format, cited for the field definitions used here: cz_type (C county or parish, Z NWS public forecast zone, M marine), magnitude in knots for wind and inches to the hundredth for hail, magnitude_type (EG, ES, MS, MG), damage_property and damage_crops formats, the unrestricted source field, begin_range, begin_azimuth and begin_location, begin and end latitude and longitude, the Enhanced Fujita ranges and the never-populated category field; read 25 August 2026. https://www.ncei.noaa.gov/pub/data/swdi/stormevents/csvfiles/Storm-Data-Bulk-csv-Format.pdf
5. NOAA National Centers for Environmental Information, Storm Events bulk CSV file directory, cited for the inventory of 77 annual event detail files, each carrying a data year and a file creation date, with creation stamps running to 19 August 2026; directory listing read 25 August 2026. https://www.ncei.noaa.gov/pub/data/swdi/stormevents/csvfiles/
6. National Weather Service Instruction 10-1605, Storm Data Preparation, issued 26 July 2021 under NWSPD 10-16, cited for the permitted event table and its county, zone and marine designators, section 2.1 on entering a continuous damage swath as a single event, section 2.7 on damage estimates and three significant digit rounding, section 2.7.2 on the 50 knot and one inch severe criteria, section 2.7.3 excluding debris clearing, firefighting and overtime from property damage, section 2.8 on encoding gusts as measured only if known and hail size in hundredths of an inch, and section 2.9.1 on the five calendar day episode rule. https://www.weather.gov/media/directives/010_pdfs/pd01016005curr.pdf
7. NOAA Storm Prediction Center, Severe Weather Database Files (Warning Coordination Meteorologist page), last updated 24 April 2026, cited for the tornado 1950 to 2025, hail 1955 to 2025 and damaging wind 1955 to 2025 files, the 2010 change in the severe hail criterion from 0.75 inch to 1.00 inch, the 2016 estimated F-scale ratings for more than 1,800 previously unrated tornadoes, the pre-1953 tornado underreporting note, the verification-use note and the statement that monetary loss information is highly suspect and should be used with caution, if at all; read 25 August 2026. https://www.spc.noaa.gov/wcm/
8. NOAA Storm Prediction Center, SPC Tornado, Hail, and Wind Database Format Specification for csv output, document last updated 7 April 2010, cited for the statement that the files represent data submitted to the Storm Data publication by NWS field offices, the field list, the magnitude conventions (F-scale, hail size in inches, wind speed in knots), the loss field changing from a damage category before 1996 to millions of dollars from 1996, the "Entry of 0 does not mean $0" warning, the ns, sn and sg tornado segment fields and the fc field marking the 1,864 records between 1953 and 1982 whose F-scale was estimated retroactively. https://www.spc.noaa.gov/wcm/data/SPC_severe_database_description.pdf
9. NOAA Storm Prediction Center, Tornado FAQ, cited for the statement that the Enhanced F scale took effect on 1 February 2007 and for the characterisation of damage rating as, at best, an exercise in educated guessing; read 25 August 2026. https://www.spc.noaa.gov/faq/tornado/
10. NOAA Storm Prediction Center, The Enhanced Fujita Scale, cited for the scale relating tornado damage to the fastest quarter-mile wind at the height of a damaged structure, the 28 damage indicators with degrees of damage carrying expected, lower bound and upper bound wind estimates, and the statement that F-scale winds are estimated from structural or tree damage; read 25 August 2026. https://www.spc.noaa.gov/efscale/
11. NOAA National Weather Service, SKYWARN, cited for the volunteer spotter programme of between 350,000 and 400,000 trained severe weather spotters, the composition of the volunteer base and the free training of about two hours; read 25 August 2026. https://www.weather.gov/skywarn/
12. NOAA National Hurricane Center, Data Archive, cited for HURDAT2: the Atlantic best track database covering 1851 to 2025 and the northeast and north central Pacific database covering 1949 to 2025, both last updated 27 February 2026, in comma-delimited text with six-hourly position, maximum winds, central pressure and, beginning in 2004, size; read 25 August 2026. https://www.nhc.noaa.gov/data/
13. American Society of Civil Engineers, About the ASCE Hazard Tool, cited for the tool being free to all users, covering wind, tornado, seismic, ice, rain, flood, snow and tsunami for ASCE/SEI 7-10, 7-16, 7-22, 41-17 and 41-23, and returning three-second gust wind speeds at 33 feet above ground for Exposure Category C by risk category; read 25 August 2026. https://www.asce.org/publications-and-news/asce-hazard-tool/about
14. T. P. McAllister, N. Wang and B. R. Ellingwood, "Risk-Informed Mean Recurrence Intervals for Updated Wind Maps in ASCE 7-16", Journal of Structural Engineering, volume 144, issue 5, article 06018001, published 2018 and cited here for the mean recurrence intervals of 300, 700, 1,700 and 3,000 years adopted for Risk Categories I to IV and the corresponding target reliability indices of 2.5, 3.0, 3.25 and 3.5. https://doi.org/10.1061/(ASCE)ST.1943-541X.0002011
15. FEMA, National Risk Index, version 1.20 of December 2025, cited for the index covering 18 hazards including hail, strong wind, tornado and hurricane with expected annual loss, social vulnerability and community resilience at county and census tract level; page read in the browser on 22 August 2026 because fema.gov refuses non-browser clients. https://www.fema.gov/flood-maps/products-tools/national-risk-index
16. MMCG Research, tabulation of the NOAA Storm Prediction Center severe weather database (damaging wind 1955 to 2025, hail 1955 to 2025 and tornado 1950 to 2025 files), computed 25 August 2026: 562,088 wind reports, 414,481 hail reports and 73,458 tornado records; magnitude-type shares and gust value distributions for 2006 to 2025; hail size distribution and sub-one-inch shares by five-year period; and tornado reports by damage rating and five-year period. https://mmcganalytics.com