Data — Reproduction package

Everything on this site can be regenerated from a single parquet file (3.1 MB). Below: download links, schemas, and example queries.

The cleaned dataset

File: tornadoes.parquet (3.1 MB; 80,626 rows; 1950–2026)

Direct download: /data/tornadoes.parquet

Schema

Column	Type	Description
`event_id`	int64	Storm Events event identifier
`episode_id`	int64 (nullable)	Parent episode identifier
`year`	int16	Calendar year
`month`	int8	Calendar month (1–12)
`begin_date`	datetime64[ns]	Beginning of event (local time)
`begin_time`	str	Beginning of event (HH:MM local time)
`state`	str (UPPERCASE)	US state name
`state_fips`	str	Two-digit FIPS code (e.g. "40")
`cz_fips`	str	County/zone FIPS code
`cz_name`	str (UPPERCASE)	County or zone name
`mag_scale`	str	F0, F1, ..., F5, EF0, ..., EF5, or empty
`mag_num`	float32	Numeric scale (0–5); null where unrated
`begin_lat`	float32	Touchdown latitude (decimal degrees, may be null)
`begin_lon`	float32	Touchdown longitude (decimal degrees, may be null)
`end_lat`	float32	Liftoff latitude (may be null)
`end_lon`	float32	Liftoff longitude (may be null)
`tor_length`	float32	Path length (miles)
`tor_width`	float32	Path width (yards)
`injuries_dir`	Int32	Direct injuries
`injuries_indir`	Int32	Indirect injuries
`deaths_dir`	Int32	Direct deaths
`deaths_indir`	Int32	Indirect deaths
`damage_prop`	str	Raw damage property (e.g. "25K", "1.5M")
`damage_crops`	str	Raw damage crops
`wfo`	str	NWS Weather Forecast Office
`source`	str	Reporting source

Quickstart (Python)

import pandas as pd
df = pd.read_parquet("tornadoes.parquet")

# Count by year
df.groupby("year").size()

# Filter to F/EF1+
sig = df[df.mag_num >= 1]

# Compute the headline centroid shift
p1_lat = sig[sig.year <= 1985].begin_lat.mean()
p1_lon = sig[sig.year <= 1985].begin_lon.mean()
p3_lat = sig[sig.year >= 2001].begin_lat.mean()
p3_lon = sig[sig.year >= 2001].begin_lon.mean()
shift_miles = 69 * ((p3_lat - p1_lat)**2 + (p3_lon - p1_lon)**2) ** 0.5
print(f"Shift: {shift_miles:.1f} mi")
# Output: Shift: 70.0 mi

Aggregated CSV files

For users who don't need the full parquet:

by_year.csv — annual counts, totals, intensities, casualties (78 rows: 1950–2024, plus 2025–2026 working rows)
by_state_year.csv — per-state per-year counts (≈4,500 rows)
by_state_era.csv — per-state P1/P2/P3 totals with per-year means and the P3:P1 ratio (54 rows)

Headline JSON

A condensed summary of every metric on this site:

/headline.json

Use this for quick checks or for driving an alternate visualization.

Source attribution

All data is from the NOAA Storm Events Database, downloaded 2026-07-12 from https://www.ncei.noaa.gov/pub/data/swdi/stormevents/csvfiles/. Bulk downloads are organized as StormEvents_details-ftp_v1.0_d<YEAR>_c<DATE>.csv.gz, one per calendar year, 1950 through present.

We do not redistribute NOAA data without attribution. The cleaned parquet file is a re-organization of the public NCEI release; it adds derived columns (year from BEGIN_DATE_TIME, month from BEGIN_DATE_TIME, mag_num from TOR_F_SCALE) but does not modify any underlying values.