saturn·

noaa significant storms noaa significant storms

source /home/coolhand/datasets/noaa-significant-storms/noaa_significant_storms.json 14,770 rows 14 columns profiled 2026-05-01 raw JSON static .html .ipynb Report Notebook

Reading

dataset summary · high confidence anthropic:claude-opus-4-7

This dataset contains 14,770 significant US storm events from the NOAA Storm Events Database, with 14 columns covering event type, location, date, magnitude, casualties, and property damage. Tornadoes dominate at 6,334 records (about 43% of rows), followed by Flash Flood, Thunderstorm Wind, and Flood — worth focusing on first since event_type drives most other fields. Geographically the events skew heavily to the central/southern US, with Texas alone accounting for 1,450 records and a long tail across 65 state values. Fatalities and injuries are highly zero-inflated (around 69% and 68% zeros respectively), so any casualty analysis should treat the non-zero tail separately. Note also that magnitude is missing for 51.8% of rows and damage_property is stored as text codes like '2.5M' and '1.00M' rather than numbers, which will need parsing before quantitative use.

citing: row_count · column_count · event_type.top_values · event_type.top_rate · state.top_values · state.n_unique · fatalities.top_rate · injuries.top_rate · magnitude.null_rate · damage_property.top_values · country.top_value · source.top_value

Schema

14 columns
Per-column summary. Click column name to jump to its detail.
Alerts
latitude numeric 0.0% 7,810
longitude numeric 0.0% 8,828
name text 0.0% 6,660
multilingual duplicates
description text 0.0% 5,796
multilingual duplicates
category categorical 0.0% 1
imbalance
date text 0.0% 5,058
one_word allcaps short_text duplicates
country categorical 0.0% 1
imbalance
event_type categorical 0.0% 17
state categorical 0.0% 65
magnitude categorical 51.8% 170
null_rate
injuries categorical 0.0% 178
fatalities categorical 0.0% 49
damage_property text 0.0% 1,014
one_word allcaps short_text duplicates
source categorical 0.0% 1
imbalance

latitude

numeric feature
Geographic latitude coordinates spanning from -14.3236 to 70.1269, with a mean of 37.28 and median of 37.12 indicating most observations cluster in the northern mid-latitudes. The tight IQR (33.63 to 41.13) suggests a heavy concentration in temperate Northern Hemisphere regions, with 159 outliers (1.08%) likely representing equatorial or far-northern points. Distribution is nearly symmetric (skew -0.18) but moderately peaked (kurtosis 3.34). Treatment: Pair with longitude for geospatial features; consider binning or clustering rather than using raw value in linear models. high · anthropic:claude-opus-4-7
n
14,770
nulls
0 (0.0%)
unique
7,810
min
-14.32
max
70.13
mean
37.28
median
37.12
std
5.247
q1
33.63
q3
41.13
iqr
7.499
skew
-0.1787
kurtosis
3.341
n_outliers
159
outlier_rate
0.01077
zero_rate
0

longitude

numeric feature
Geographic longitude coordinates spanning -170.73 to 171.47, with values concentrated in the Western Hemisphere (median -90.22, IQR -96.4 to -84.23) consistent with North American locations. The distribution is heavy-tailed (kurtosis 55.6, skew 1.29) with 623 outliers (4.2%) likely representing locations outside the dominant cluster. No nulls or zeros, and 8828 unique values across 14770 rows suggests repeated locations. Treatment: Pair with latitude for geospatial features; consider clustering or binning by region rather than using raw values linearly. high · anthropic:claude-opus-4-7
n
14,770
nulls
0 (0.0%)
unique
8,828
min
-170.7
max
171.5
mean
-90.94
median
-90.22
std
11.7
q1
-96.4
q3
-84.23
iqr
12.17
skew
1.286
kurtosis
55.61
n_outliers
623
outlier_rate
0.04218
zero_rate
0

name

text label multilingual duplicates
Templated event labels of the form ' in , ' describing US severe weather incidents (tornado, flood, hail, thunderstorm wind dominate top_words). With 14,770 rows but only 6,660 unique values and a 54.9% duplicate rate, the same state/county/event combinations recur heavily — 'Hail in TEXAS, TARRANT' alone appears 59 times. The 'multilingual' alert is misleading: 4,796 strings tag as English against tiny counts in other languages, almost certainly false positives from the proper-noun template. Treatment: Parse into separate event_type, state, and county fields rather than using the raw string. high · anthropic:claude-opus-4-7
n
14,770
nulls
0 (0.0%)
unique
6,660
len_min
17
len_max
134
len_mean
30.22
len_median
29
len_p95
41
word_mean
4.588
word_median
4
n_empty
0
n_duplicates
8,110
duplicate_rate
0.5491
vocab_size
1,980
readability_flesch_mean
31.16
emoji_rate
0
url_rate
0
one_word_rate
0
allcaps_rate
0
boilerplate_rate
0

description

text metadata multilingual duplicates
This appears to be a templated event-summary field describing storm or disaster impacts (magnitude, injuries, fatalities, property damage in dollars), not free-form prose. Despite 14,770 rows, only 5,796 are unique and 60.8% are duplicates — the top value alone repeats 1,055 times — so the field carries far less information than its size suggests. The 'multilingual' alert is misleading: 4,984 rows tag as English against only 16 in other languages, likely noise from short numeric strings. Low Flesch (29.86) and a 7.4-word mean confirm terse, formulaic content rather than narrative text. Treatment: Parse out structured fields (magnitude, injuries, fatalities, damage_usd) with regex rather than embedding the raw string. high · anthropic:claude-opus-4-7
n
14,770
nulls
0 (0.0%)
unique
5,796
len_min
3
len_max
259
len_mean
50.09
len_median
36
len_p95
166
word_mean
7.393
word_median
5
n_empty
0
n_duplicates
8,974
duplicate_rate
0.6076
vocab_size
4,289
readability_flesch_mean
29.86
emoji_rate
0
url_rate
0
one_word_rate
0.0002708
allcaps_rate
0.0002708
boilerplate_rate
0

category

categorical metadata imbalance
This column is a constant tag identifying the dataset partition: every one of the 14,770 rows holds the single value "significant_us_storms". Cardinality is 1 with entropy 0.0 and a top_rate of 1.0, so it carries no information for modelling. Treatment: Drop before modelling; retain only as a provenance label. high · anthropic:claude-opus-4-7
n
14,770
nulls
0 (0.0%)
unique
1
top_value
significant_us_storms
top_rate
1
cardinality
1
entropy
0
entropy_ratio
0

date

text timestamp one_word allcaps short_text duplicates
This is a date column stored as ISO-formatted text (YYYY-MM-DD), with every value exactly 10 characters long across 14,770 rows and no nulls. Values span at least 1965 to 2021, but heavy clustering — 9,712 duplicates (65.8%) and spikes like 1974-04-03 (126 rows) and 2011-04-27 (105 rows) — suggests events grouped on shared dates rather than unique daily records. Only 5,058 distinct dates appear, so this won't act as a row identifier. Treatment: parse to datetime and use for temporal joins or time-based features. high · anthropic:claude-opus-4-7
n
14,770
nulls
0 (0.0%)
unique
5,058
len_min
10
len_max
10
len_mean
10
len_median
10
len_p95
10
word_mean
1
word_median
1
n_empty
0
n_duplicates
9,712
duplicate_rate
0.6575
vocab_size
5,058
readability_flesch_mean
121.2
emoji_rate
0
url_rate
0
one_word_rate
1
allcaps_rate
1
boilerplate_rate
0

country

categorical metadata imbalance
This column records country of origin but contains a single value, "USA", across all 14770 rows. With cardinality of 1, entropy of 0, and a top_rate of 1.0, it carries no information. The imbalance alert is effectively a constant-column flag. Treatment: Drop; constant column with no predictive signal. high · anthropic:claude-opus-4-7
n
14,770
nulls
0 (0.0%)
unique
1
top_value
USA
top_rate
1
cardinality
1
entropy
0
entropy_ratio
0

event_type

categorical label
Categorical label of severe weather event types across 14,770 rows with no nulls and only 17 distinct categories. Tornado dominates at 42.9% (6,334 records), followed by Flash Flood, Thunderstorm Wind, Flood, and Hail; tail categories like Marine Thunderstorm Wind have just 25 records. Entropy ratio of 0.57 confirms the distribution is heavily skewed toward a few classes. Treatment: One-hot or target-encode; consider grouping rare marine categories before modelling. high · anthropic:claude-opus-4-7
n
14,770
nulls
0 (0.0%)
unique
17
top_value
Tornado
top_rate
0.4288
cardinality
17
entropy
2.336
entropy_ratio
0.5715

state

categorical feature
U.S. state names stored as uppercase strings, fully populated across 14,770 rows with no nulls. Cardinality is 65, well above the 50 states, suggesting territories, districts, or non-state codes are mixed in. Distribution is broad (entropy ratio 0.86) with Texas leading at 9.8% (1,450 rows), followed by Missouri, Arkansas, Mississippi, and Georgia. Treatment: Normalize to a known state/territory code list, then one-hot or target-encode for modelling. high · anthropic:claude-opus-4-7
n
14,770
nulls
0 (0.0%)
unique
65
top_value
TEXAS
top_rate
0.09817
cardinality
65
entropy
5.182
entropy_ratio
0.8605

magnitude

categorical feature null_rate
Likely a magnitude or measurement value stored as text rather than numeric, with 170 distinct string values dominated by "0" (54.2% of non-nulls, 3863 rows). More than half the column is missing (null_rate 0.5178), and the remaining values mix small decimals like "1.75" and "2.75" with much larger ones like "70.00" and "61.00", suggesting either heterogeneous units or a compressed scale. Entropy ratio of 0.48 confirms heavy concentration on the zero bucket. Treatment: Cast to numeric, treat "0" and nulls explicitly, and investigate whether the large vs small values reflect different units before modelling. medium · anthropic:claude-opus-4-7
n
14,770
nulls
7,648 (51.8%)
unique
170
top_value
0
top_rate
0.5424
cardinality
170
entropy
3.586
entropy_ratio
0.484

injuries

categorical feature
This is an injury count stored as strings, with 178 distinct values dominated by '0' (68.1% of 14,770 rows). The next most common values ('1' through '7', plus '10' and '12') are clearly numeric, suggesting the column should be cast to integer rather than treated as categorical. Low entropy ratio (0.33) reflects the heavy zero-mass. Treatment: Cast to integer and consider log1p or zero-inflated treatment given the zero-heavy distribution. high · anthropic:claude-opus-4-7
n
14,770
nulls
0 (0.0%)
unique
178
top_value
0
top_rate
0.6814
cardinality
178
entropy
2.468
entropy_ratio
0.3301

fatalities

categorical numeric_target
Counts of fatalities per event, stored as strings but numeric in content with 49 distinct values. Heavily zero-inflated: 69.1% of 14,770 rows are "0" and the next bucket "1" covers 3,208 more, leaving a long thin tail (5+ fatalities each appear in under 100 rows). Low entropy ratio (0.25) confirms the distribution is dominated by a single value. Treatment: Cast to integer and model as a zero-inflated count (e.g., ZIP/NB) or binarise to fatal/non-fatal. high · anthropic:claude-opus-4-7
n
14,770
nulls
0 (0.0%)
unique
49
top_value
0
top_rate
0.6912
cardinality
49
entropy
1.423
entropy_ratio
0.2535

damage_property

text feature one_word allcaps short_text duplicates
This column encodes property damage estimates as short magnitude-suffixed strings (e.g. '2.5M', '250K', '0.00K'), with every value being a single token of at most 8 characters. The format is inconsistent — some values use two decimals ('1.00M') while others don't ('1M', '25M') — and 368 rows are empty strings rather than nulls, with the literal '0.00K' appearing 1229 times to denote zero damage. Duplication is extreme (93.1%) because the underlying domain is a small set of round-number estimates, yielding only 1014 distinct values. Treatment: Parse the K/M/B suffix into a numeric float, treat empty strings and '0.00K' explicitly, then log-transform before modelling. high · anthropic:claude-opus-4-7
n
14,770
nulls
0 (0.0%)
unique
1,014
len_min
0
len_max
8
len_mean
4.381
len_median
5
len_p95
7
word_mean
1
word_median
1
n_empty
368
n_duplicates
13,756
duplicate_rate
0.9313
vocab_size
1,013
readability_flesch_mean
117
emoji_rate
0
url_rate
0
one_word_rate
1
allcaps_rate
0.8724
boilerplate_rate
0

source

categorical metadata imbalance
This column records the dataset's provenance, with every one of the 14,770 rows tagged 'NOAA Storm Events Database'. Cardinality is 1 and entropy is 0, so it carries no discriminative signal whatsoever. Treatment: Drop before modelling; retain only as dataset-level provenance. high · anthropic:claude-opus-4-7
n
14,770
nulls
0 (0.0%)
unique
1
top_value
NOAA Storm Events Database
top_rate
1
cardinality
1
entropy
0
entropy_ratio
0