NOAA Storm Events County Peril Climatology

Turns NOAA/NCEI's Storm Events Database into a per-county, per-peril climatology: how often each severe-weather peril has hit each county, what it cost, who it killed or injured, and how bad the worst year was — in one row per (county x peril).

Built directly from NCEI's authoritative bulk .csv.gz archive (1950 to the current year), not a scraped front-end. No API key.

What one row is

One county, one peril type, over the year range you asked for. For example "Yuma County, Colorado — Thunderstorm Wind, 2023-2024: 55 events across 23 distinct storm episodes, max gust 87 kt, $147,140 in reported damage, peril score 54.18."

Read this before you use it: county vs forecast zone

NOAA records some perils against counties (CZ_TYPE = "C") and others against NWS forecast zones (CZ_TYPE = "Z"). This is not a minor detail: 43% of all national Storm Events records are zone-typed, and across the full 2023, 2024 and 2025 national files 37 of NOAA's 54 peril types have ZERO county-typed events, with seven more -- Wildfire, Drought, Heat, Heavy Snow, High Wind, Strong Wind and Winter Storm -- recording a handful of county events against thousands of zone events (Drought: 1 county vs 12,570 zone).

Among the perils a county-only run cannot see:

Wildfire · Excessive Heat · Heat · Winter Storm · Heavy Snow · Blizzard · Ice Storm · Drought · Hurricane (Typhoon) · Tropical Storm · Storm Surge/Tide · Coastal Flood · High Wind · Strong Wind · Dense Fog · Dense Smoke · Extreme Cold/Wind Chill · Avalanche · Rip Current

includeZones is off by default, which keeps every row tied to a real county FIPS. When it is off:

• every output row reports zone_events_excluded (how many in-scope events were dropped) and excluded_zone_perils (which perils were dropped by name);
• if you filter eventTypes to perils NOAA only zone-types, the run fails with an actionable message instead of handing you an empty dataset.

Set "includeZones": true to include them; those rows carry geography_basis: "forecast_zone" and are keyed to the zone's FIPS.

Who this is for

• Property & casualty underwriters — hail, wind and tornado frequency and severity per county, with max hail diameter and max gust.
• CRE / infrastructure siting — a peril profile for a candidate county before committing capital.
• Cat modellers and reinsurance analysts — a free, reproducible baseline built from the authoritative NCEI archive.
• Climate and resilience researchers — county-level event, casualty and loss history back to 1950.
• AI agents — one call per state, one row per (county x peril), with explicit completeness fields so the agent knows what it does and does not know.

Example input

{
  "states": ["Colorado"],
  "counties": [],
  "startYear": 2023,
  "endYear": 2025,
  "eventTypes": [],
  "includeZones": false,
  "maxResults": 2000
}

counties accepts either a 5-digit FIPS ("08031") or a bare county name matched case-insensitively ("Denver"). states accepts "CO" or "Colorado", as an array or a comma-separated string.

Output — every field

Identity

Field	Meaning
county_fips	5-digit county FIPS, built as pad2(STATE_FIPS) + pad3(CZ_FIPS).
county_name	NOAA's CZ_NAME.
state	State name in NOAA's uppercase form.
state_abbr	Two-letter postal code.
event_type	The NOAA peril type.
geography_basis	county (CZ_TYPE=C), forecast_zone (CZ_TYPE=Z), or mixed.

Frequency and severity

Field	Meaning
event_count	NOAA event records in scope.
episode_count	Distinct NOAA storm episodes behind them. One storm system can log many events in one county, so this is the better "how often did a distinct storm hit here" measure.
events_per_year	event_count divided by n_years.
deaths, injuries, casualties	Combined totals.
deaths_direct, deaths_indirect, injuries_direct, injuries_indirect	NOAA's own direct/indirect split.
worst_year	Year with the highest combined damage.
worst_year_damage_usd	Damage in that year.
first_event_date, last_event_date	Earliest and latest event begin date in scope.
max_magnitude	Largest NOAA magnitude - hail diameter in inches, or wind gust in knots.
magnitude_type	What max_magnitude measures: EG/MG (gust), MS/ES (sustained), or empty for hail.
max_tornado_ef_scale	Strongest EF/F rating recorded.
max_tornado_length_miles, max_tornado_width_yards	Largest tornado path recorded.
flood_causes	Distinct NOAA flood causes.
report_sources	Who reported the events (Trained Spotter, Emergency Manager, ASOS, ...).
nws_offices	NWS Weather Forecast Offices that logged them.
peril_score	EDITORIAL 0-100 composite: annualized frequency (0.40), casualties (0.25), total damage (0.35), each log-saturating. A triage heuristic, not an actuarial loss estimate.

Damage — and how complete it is

NOAA leaves DAMAGE_PROPERTY empty on 21.6% of events (2024 national file): no dollar figure was ever estimated. That is not the same as $0.

Field	Meaning
property_damage_usd	Sum of reported property damage. null when NOAA costed none of this row's events.
crop_damage_usd	Same, for crops.
total_damage_usd	Property + crop. null when neither was ever reported.
damage_per_year_usd	total_damage_usd divided by n_years, or null.
events_with_damage_reported	How many events carry a NOAA dollar figure.
events_missing_damage_report	How many do not.
damage_report_completeness_pct	The percentage. 0 means the damage columns are null.
damage_is_lower_bound	true when some but not all events were costed, so the sum is a floor.

peril_score deliberately treats an unreported figure as $0 so scores stay comparable across rows, even where the damage columns are null.

Coverage and provenance — read these before trusting a row

Field	Meaning
years_covered	The years this row is actually built from.
years_requested	The range you asked for.
years_retrieved	Comma-separated list of years whose NCEI file was downloaded and parsed.
years_missing	Requested years that were absent from the index or failed to download. null when nothing was missing.
data_complete	true only when every requested year was retrieved.
source_status	ok or partial.
n_years	Denominator for the per-year metrics - the count of retrieved years, not requested ones.
zone_events_excluded	In-scope events dropped because they are zone-typed and includeZones is off.
excluded_zone_perils	The peril types dropped for that reason, by name.
source_files	The exact NCEI .csv.gz filenames used.
source_url	The NCEI directory.

How it works

1. Fetches the NCEI csvfiles/ directory index and asserts it live: it must still list at least 60 years back to 1960 or earlier, or the run fails before anything is billed. The year bounds you can request are derived from that index rather than hard-coded.
2. Resolves each year's StormEvents_details-ftp_v1.0_d{YEAR}_c{DATE}.csv.gz by the newest creation date, and logs which one it used.
3. Downloads and gunzips each annual file, with retry and backoff, and stream-filters it to your states while parsing - the full ~75,000-row national array is never materialised, because building it exhausts a 4 GB run.
4. Asserts the file's shape live: 25 required columns present, the national event count inside a measured band, and CZ_TYPE still a two-value vocabulary. Any of those failing fails the run and bills nothing. A previously-unseen EVENT_TYPE is reported as a warning, not a failure.
5. Aggregates to one row per (county x peril), scores, sorts by score, and applies maxResults.
6. Gates before writing: if no year could be retrieved, if the state matched nothing, or if the filters matched no rows, the run fails with a message naming the cause - nothing is emitted and nothing is billed.

Use as an MCP tool

Available through mcp.apify.com. One call per state; one row per (county x peril). An agent should read data_complete, source_status and damage_report_completeness_pct before quoting a number, and read excluded_zone_perils before concluding a county has no history of a peril.

Pricing

Pay-per-result: one dataset row = one billed result. A failed run emits nothing and costs nothing beyond the platform's Actor-start event. Volume tiers apply automatically on paid Apify plans.

FAQ

Why did I get no Wildfire / Excessive Heat / Winter Storm rows? Because NOAA records those against NWS forecast zones, not counties, and includeZones is off by default. Set it to true. Your rows already name the exclusions in excluded_zone_perils.

Why is total_damage_usd null on some rows? Because NOAA published no dollar estimate for any event in that row. Publishing $0 there would be a factual claim the source never made. Check damage_report_completeness_pct.

Why does years_covered differ from what I asked for? Because at least one annual file could not be retrieved. years_missing names them and data_complete is false.

Is peril_score an actuarial number? No. It is a transparent, documented, absolute composite for triage and ranking. The inputs are all on the row, so you can build your own.

How far back does it go? The NCEI archive starts in 1950, but pre-1996 files only carry tornado, thunderstorm-wind and hail.

Related actors

• fema-nri-county-risk-profile - FEMA's National Risk Index for the same county, at county and tract resolution.
• wildfire-asset-exposure-screener - live and historical wildfire perimeters and Wildfire Hazard Potential for a coordinate.
• nfip-flood-loss-risk-screener - paid NFIP flood-claim history near a point.