How to Read NOAA's Storm Events Database in Python

On this page

Problem statement

NOAA's Storm Events Database is the official US record of damaging and notable weather: every tornado, hail report, flash flood, heat wave and blizzard that a National Weather Service office documented, with times, places, deaths, injuries and damage estimates. The bulk CSV for 2025 has 72,360 rows and 51 columns, and it is not ready to analyse as it arrives. Times are local to each county with a two-digit year, two-fifths of the rows have no coordinates because they describe a forecast zone rather than a point, and damage is stored as text such as 10.00K. This page reads the 2025 file into pandas, converts the times to UTC, builds a point layer from the events that have positions, and compares 2025 with 2024.

Quick answer

Download the year's StormEvents_details file from NCEI's bulk CSV directory and read it with pd.read_csv(..., low_memory=False). Then parse BEGIN_DATE_TIME with format="%d-%b-%y %H:%M:%S", subtract the hour offset in CZ_TIMEZONE to get UTC, convert the damage strings with their K/M/B multipliers, and build points from BEGIN_LAT/BEGIN_LON where they exist. The setup's times() and dollars() functions do all of that.

raw = pd.read_csv("data/storm-events-details-2025-c20260819.csv.gz", low_memory=False)
print(f"{len(raw):,} rows x {raw.shape[1]} columns")
print(raw.EVENT_TYPE.value_counts().head(6).to_string())

72,360 rows x 51 columns
EVENT_TYPE
Thunderstorm Wind    21807
Hail                  9205
Flash Flood           5393
High Wind             4603
Winter Weather        4436
Drought               3283

Thunderstorm wind is the most common event type, with 21,807 reports, ahead of hail (9,205) and flash floods (5,393). The list also includes drought and winter weather, which are reported by forecast zone and last for days or months.

Map of the lower 48 states covered in coloured dots for 2025 Storm Events with a begin point: thunderstorm wind in orange across the East and Plains, hail in blue concentrated on the Plains, flash floods in dark blue, tornadoes in red in the South and Midwest, and other types in grey. — Every 2025 Storm Events record with a begin point, by type. Convective weather dominates because heat, snow and wind rows are reported by zone and have no coordinates.

Files, times, positions and damage

1. The files

fat = pd.read_csv("data/storm-events-fatalities-2025-c20260819.csv.gz")
print(f"details: {len(ev):,} events; fatalities: {len(fat):,} deaths, joined on EVENT_ID "
      f"({fat.EVENT_ID.isin(ev.EVENT_ID).mean():.0%} found)")
print("fields per event:", ", ".join(ev.columns[:12]), "...")

details: 72,360 events; fatalities: 895 deaths, joined on EVENT_ID (100% found)
fields per event: BEGIN_YEARMONTH, BEGIN_DAY, BEGIN_TIME, END_YEARMONTH, END_DAY, END_TIME, EPISODE_ID, EVENT_ID, STATE, STATE_FIPS, YEAR, MONTH_NAME ...

NCEI publishes three files per year. The details file holds one row per event. The fatalities file holds one row per death, with age, sex and where the person was, and every one of its 895 rows joins to a detail row through EVENT_ID. The locations file holds extra points along some events' paths. The details file is the one most analyses need.

2. Times are local, with a two-digit year

print(ev[["BEGIN_DATE_TIME", "CZ_TIMEZONE", "begin", "begin_utc"]].head(3).to_string(index=False))
print(ev.CZ_TIMEZONE.value_counts().to_dict())

   BEGIN_DATE_TIME CZ_TIMEZONE               begin           begin_utc
31-MAR-25 11:04:00       EST-5 2025-03-31 11:04:00 2025-03-31 16:04:00
30-MAR-25 15:52:00       EST-5 2025-03-30 15:52:00 2025-03-30 20:52:00
05-JAN-25 18:00:00       EST-5 2025-01-05 18:00:00 2025-01-05 23:00:00
{'CST-6': 35400, 'EST-5': 24782, 'MST-7': 8601, 'PST-8': 2578, 'AST-4': 454, 'AKST-9': 253, 'HST-10': 126, 'SST-11': 93, 'GST10': 34, 'PDT-7': 30, 'EDT-4': 9}

31-MAR-25 11:04:00 is local standard time in the zone named beside it. EST-5 means five hours behind UTC, so the event began at 16:04 UTC. Eleven different zone labels appear, from SST-11 in American Samoa to GST10 on Guam, which is ten hours ahead of UTC and has no minus sign. A small number of rows are labelled with daylight time (PDT-7, EDT-4). Sort or bin by begin_utc when comparing places, and by the local begin for the hour of day.

3. Some events have coordinates, some only a zone

has_xy = ev.BEGIN_LAT.notna()
print(f"events with a begin point: {has_xy.mean():.1%}")
print(ev.groupby("CZ_TYPE").BEGIN_LAT.apply(lambda s: f"{s.notna().mean():.1%} with coordinates of {len(s):,}").to_string())
pts = gpd.GeoDataFrame(ev[has_xy], geometry=gpd.points_from_xy(ev.BEGIN_LON[has_xy], ev.BEGIN_LAT[has_xy]), crs=4326)
print(f"point layer: {len(pts):,} events")

events with a begin point: 60.7%
CZ_TYPE
C    99.7% with coordinates of 41,650
Z     7.8% with coordinates of 30,710
point layer: 43,917 events

60.7% of events have a begin point. The split follows CZ_TYPE: 99.7% of county-based events (C: hail, tornadoes, thunderstorm wind, floods) carry coordinates, against 7.8% of zone-based ones (Z: heat, winter weather, high wind, drought). A point map of storm events is therefore a map of convective weather. Heat and snow appear only if you join the zone rows to zone boundaries.

Horizontal bar chart of the 14 most common event types, each bar showing how many events have a begin point: thunderstorm wind, hail, flash flood, flood, marine thunderstorm wind and tornado are almost fully coloured, while high wind, winter weather, drought, winter storm, heat, heavy snow, cold and excessive heat, all marked zone, are almost entirely uncoloured. — Whether an event has coordinates depends on its type: county-based types carry a begin point, zone-based ones do not (99.7% against 7.8%).

4. Damage is text

print(ev.DAMAGE_PROPERTY.dropna().sample(5, random_state=1).tolist())
print(f"property damage reported: ${ev.prop.sum() / 1e9:.2f} billion; crops ${ev.crops.sum() / 1e9:.2f} billion; "
      f"blank property damage: {ev.DAMAGE_PROPERTY.isna().mean():.1%}")

['1.00K', '0.00K', '1.00K', '10.00K', '8.00K']
property damage reported: $4.50 billion; crops $0.27 billion; blank property damage: 22.7%

Damage arrives as strings such as 1.00K and must be parsed with its multiplier. Parsed, 2025 property damage sums to $4.50 billion and crop damage to $0.27 billion. 22.7% of rows have no property estimate at all, which is not the same as zero.

States, years and a point layer

The data used on this page

The National Centers for Environmental Information's Storm Events Database bulk CSV files for 2025 and 2024.

from pathlib import Path

import geopandas as gpd
import numpy as np
import pandas as pd
import requests

Path("data").mkdir(exist_ok=True)
# NOAA NCEI Storm Events Database: the 2025 details and fatalities files (revision c20260819) and the 2024 details file
# (revision c20260728), saved on this site on 15 October 2026 because NCEI replaces a year's files whenever it revises
# them and the old names disappear. Live: https://www.ncei.noaa.gov/pub/data/swdi/stormevents/csvfiles/
FILES = ["storm-events-details-2025-c20260819.csv.gz", "storm-events-fatalities-2025-c20260819.csv.gz",
         "storm-events-details-2024-c20260728.csv.gz"]
for name in FILES:
    if not Path(f"data/{name}").exists():
        Path(f"data/{name}").write_bytes(requests.get(f"https://data.spatialworkflow.io/{name}", timeout=600).content)
ev = pd.read_csv("data/storm-events-details-2025-c20260819.csv.gz", low_memory=False)


def dollars(s):
    """Damage is stored as text: '10.00K', '1.60B', '0' or blank. A blank means no estimate, not zero."""
    s = s.str.strip()
    return pd.to_numeric(s.str.rstrip("KMB"), errors="coerce") * s.str[-1].map({"K": 1e3, "M": 1e6, "B": 1e9}).fillna(1.0)


def times(df):
    """Event times are local to the county or zone ('CST-6'), with a two-digit year."""
    df["begin"] = pd.to_datetime(df.BEGIN_DATE_TIME, format="%d-%b-%y %H:%M:%S")
    df["end"] = pd.to_datetime(df.END_DATE_TIME, format="%d-%b-%y %H:%M:%S")
    df["utc_offset"] = df.CZ_TIMEZONE.str.extract(r"(-?\d+)$")[0].astype(int)
    df["begin_utc"] = df.begin - pd.to_timedelta(df.utc_offset, unit="h")
    df["prop"], df["crops"] = dollars(df.DAMAGE_PROPERTY), dollars(df.DAMAGE_CROPS)
    return df


ev = times(ev)
print(f"{len(ev):,} Storm Events records for 2025, {ev.EPISODE_ID.nunique():,} episodes, {ev.EVENT_TYPE.nunique()} event types")

72,360 Storm Events records for 2025, 9,846 episodes, 48 event types

Example 1: events by state

by_state = ev.groupby("STATE").agg(events=("EVENT_ID", "size"), deaths=("DEATHS_DIRECT", "sum"),
                                   damage_musd=("prop", lambda s: s.sum() / 1e6)).sort_values("events", ascending=False)
print(by_state.head(8).round(1).to_string())

              events  deaths  damage_musd
STATE                                    
TEXAS           5641     153        488.5
OKLAHOMA        2892      13         87.6
MISSOURI        2757      21       1765.2
ILLINOIS        2702       9        532.5
VIRGINIA        2632       7         24.8
SOUTH DAKOTA    2400       3          1.2
TENNESSEE       2343      17         53.6
ALABAMA         2302      11         21.8

Texas reported the most events (5,641) and the most direct deaths (153), most of them in the July floods. Missouri's $1,765.2 million of damage is almost all one event: the St. Louis tornado of 16 May. Report counts also reflect how many people are around to report: Virginia's 2,632 events produced $24.8 million of damage, and South Dakota's 2,400 produced $1.2 million.

Example 2: 2025 against 2024

ev24 = times(pd.read_csv("data/storm-events-details-2024-c20260728.csv.gz", low_memory=False))
cmp = pd.DataFrame({"2024": ev24.EVENT_TYPE.value_counts(), "2025": ev.EVENT_TYPE.value_counts()}).fillna(0).astype(int)
cmp["change"] = (cmp["2025"] / cmp["2024"] - 1).round(3)
print(cmp.sort_values("2025", ascending=False).head(8).to_string())

                    2024   2025  change
EVENT_TYPE                             
Thunderstorm Wind  19940  21807   0.094
Hail                8941   9205   0.030
Flash Flood         4661   5393   0.157
High Wind           3994   4603   0.152
Winter Weather      3557   4436   0.247
Drought             3473   3283  -0.055
Winter Storm        2112   3070   0.454
Heat                3171   2864  -0.097

Year-on-year changes are large for the zone-based types. Winter storm events rose 45.4% from 2024 and heat events fell 9.7%. These are counts of county and zone segments, not of separate storms: one storm crossing ten zones is ten rows. The concept page on episodes covers what that does to counts.

Dot-and-line chart on a log scale comparing 2024 and 2025 event counts for the 12 most common types; most move little, winter storm and winter weather rise, and extreme cold/wind chill rises steeply from a low 2024 count. — Records by type in 2024 and 2025. The big swings are in zone-based winter and cold types, where one storm becomes many zone rows.

Example 3: a clean point layer to keep

keep = pts[["EVENT_ID", "EPISODE_ID", "STATE", "EVENT_TYPE", "begin_utc", "MAGNITUDE", "DEATHS_DIRECT", "prop", "geometry"]]
keep.to_file("storm_events_2025.gpkg", layer="events", driver="GPKG")
print(f"wrote {len(keep):,} points, {Path('storm_events_2025.gpkg').stat().st_size / 1e6:.1f} MB")

wrote 43,917 points, 6.7 MB

The 43,917 positioned events fit in a 6.7 MB GeoPackage with UTC times and parsed damage, ready for QGIS or a web map. Keep EVENT_ID and EPISODE_ID so you can join back to the narratives and to the fatalities file.

Where the records come from

Each Weather Service forecast office compiles Storm Data for its own area every month, from spotter reports, emergency managers, media, damage surveys and its own radar analysis. NCEI publishes the result. It is a record of what offices documented, not an instrument record, which shows in three ways. Rural areas report less, and a quarter-sized hailstone in an empty county never becomes a row. Damage figures are office estimates, often rounded or left blank. And revisions keep arriving: the 2025 file was reissued in August 2026 (c20260819), and NCEI deletes the older revision when it does. That is why this page reads a dated copy saved on this site.

Things that trip up a first read

The two-digit year. %y maps 69–99 to the 1900s and 00–68 to the 2000s, so in the database's early files a 55 becomes the 2050s. For files from before 1969, take the year from the YEAR column or BEGIN_YEARMONTH.
Marine and lake rows. STATE is sometimes GULF OF MEXICO or LAKE MICHIGAN, with marine-zone codes that are not in any land boundary file.
Episode and event narratives are free text, sometimes thousands of characters long. Leave them out of a GeoPackage you intend to tile.
The current year's file grows every month until about five months after the year ends, so record which revision you read.

Internal links

Storm Events, episodes and county segments — what one row is
Storm Events damage columns will not sum — the damage strings
Storm Events will not join to counties — county and zone codes
How to map tornado tracks — the begin and end points as lines

FAQ

Where do I download the NOAA Storm Events database?

The bulk CSV files are in NCEI's pub/data/swdi/stormevents/csvfiles/ directory, one details, fatalities and locations file per year, gzipped. The file name carries a revision date that changes when the file is reissued.

Are Storm Events times in UTC?

No. BEGIN_DATE_TIME is local standard time of the county or zone, named in CZ_TIMEZONE (for example CST-6). Subtract the offset to get UTC.