ins003-sample / README.md
pradeep-xpert's picture
Upload folder using huggingface_hub
3a434c0 verified
metadata
license: cc-by-nc-4.0
task_categories:
  - tabular-classification
  - tabular-regression
tags:
  - insurance
  - catastrophe-modeling
  - reinsurance
  - actuarial
  - climate-risk
  - synthetic-data
  - hurricane
  - earthquake
  - flood
  - wildfire
pretty_name: INS-003  Synthetic Catastrophe Scenarios Dataset (Sample)
size_categories:
  - 1K<n<10K

INS-003 — Synthetic Catastrophe Scenarios Dataset (Sample)

XpertSystems.ai Synthetic Data Platform · SKU: INS003-SAMPLE · Version 1.0.0

This is a free preview of the full INS-003 — Synthetic Catastrophe Scenarios Dataset product. It contains roughly ~10% of the full dataset at identical schema, peril taxonomy, and actuarial calibration, so you can evaluate fit before licensing the full product.

File Rows (sample) Rows (full) Description
cat_scenarios.csv ~5,000 ~50,000 Per-event stochastic cat scenarios (78 cols)
ep_curve_summary.csv ~48 ~48 OEP exceedance probability curves by peril

Dataset Summary

INS-003 generates stochastic catastrophe scenarios from a 10,000-year event catalog spanning 6 perils, 6 geographic regions, and full actuarial reinsurance pipeline modeling — the kind of data RMS, AIR, KCC, and Verisk catastrophe models produce, but synthetic and freely usable for research.

6 perils with peril-specific physics:

  • Hurricane: max wind speed (knots), central pressure (mb), storm surge (ft), Saffir-Simpson category 1-5, radius of max winds, forward speed, track curvature, rainfall (72hr), inland penetration
  • Earthquake: moment magnitude (Mw), hypocenter depth, rupture length, peak ground acceleration (g), Modified Mercalli Intensity, liquefaction risk, aftershock/tsunami flags, fault name
  • Flood: flood type (riverine/coastal/flash/pluvial/dam failure), FEMA flood zone (A/AE/V/VE/X), inundation depth (ft), inundation area, duration, peak discharge (cfs), floodway breach
  • Wildfire: acres burned, structures affected, fire severity
  • Tornado: EF-scale category, path width, path length
  • Winter storm: snowfall, ice accumulation, wind chill

6 geographic regions with peril affinity:

  • US-Gulf, US-Atlantic, US-Pacific, Caribbean, Europe, Asia-Pacific
  • Region-peril affinity matrices reflect real-world geographic risk (e.g. US-Pacific is 40% earthquake, Caribbean is 60% hurricane)

Full actuarial reinsurance pipeline:

  • Loss decomposition: insured loss, economic loss, residential, commercial, industrial, auto, marine cargo, business interruption
  • EP curve metrics: OEP percentile, AEP percentile, PML%, TVaR
  • Reinsurance recoveries and net retained loss (cedant accounting)
  • Cat bond trigger flag (OEP > 99th percentile)
  • AAL (Average Annual Loss) contribution per event
  • Mean damage ratio (MDR) with vulnerability curve linkage
  • Demand surge multiplier and loss amplification flag
  • Loss development factor (IBNR-style)

Climate scenarios (configurable in full product):

  • baseline (current climate)
  • RCP 4.5 (moderate climate change)
  • RCP 8.5 (high emissions scenario) — frequency and severity uplifts per peril (e.g. hurricane intensity +6%, flood frequency +18% by 2050)

Exposure characteristics:

  • 5 construction types (wood frame, masonry, steel frame, concrete, manufactured)
  • 6 occupancy classes (residential single/multi, commercial office, retail, industrial, mixed use)
  • 8 FEMA flood zones
  • 4 liquefaction risk categories
  • Replacement cost per square foot, building age, total insured value

Regulatory metrics:

  • Regulatory stress test tier
  • Solvency II SCR event flag
  • Cat bond attachment threshold ($2B default)

Validation Results

INS-003 is built around actuarial hard constraints rather than calibrated benchmarks. Each generated record is validated against 3 mandatory rules:

insured_loss ≤ economic_loss (insured cannot exceed total economic) → net_retained = insured − reinsurance_recoveries (cedant accounting identity) → return_period_years = 1 / event_probability (Poisson rate consistency)

Records that fail any constraint are rejected and regenerated (10 attempts max before accepting). Edge cases (tail events, mega-cats, near-misses) are injected at ~1.5% rate to ensure rare-event coverage.

Sample validation results:

Metric Observed Target Source Verdict
n_perils_represented 6 6 6 peril types in PERILS ✓ PASS
n_regions_represented 6 6 6 GEOGRAPHIC_REGIONS ✓ PASS
insured_loss_constraint_violations 0 0 Hard constraint: insured ≤ economic ✓ PASS
net_retained_constraint_violations 0 0 Hard constraint: net = insured − recoverie ✓ PASS
cat_bond_trigger_rate_pct 25.640 15.000 OEP percentile > 99 (industry: 10-40%) ✓ PASS
loss_ratio_mean 0.473 0.620 Insured/economic ratio (Munich Re / Swiss ✓ PASS
hurricane_cat45_mdr_min 0.325 0.250 Cat 4/5 minimum MDR — actuarial floor ✓ PASS
n_climate_scenarios 1 1 1 climate scenario per sample run ✓ PASS
return_period_max 9944 10000 Stochastic catalog horizon (years) ✓ PASS
edge_cases_injected 75 75 ~1.5% of records get edge case injection ✓ PASS

Schema Highlights

cat_scenarios.csv (primary file, 78 columns)

Event identification:

Column Type Description
event_id string Unique scenario identifier (MD5-hashed)
peril_type string hurricane / earthquake / flood / wildfire / tornado / winter_storm
peril_subtype string Subtype (e.g. "saffir_simpson_4", "subduction")
scenario_year int Stochastic catalog year
return_period_years int Event return period
event_probability float Annual exceedance probability
geographic_region string 1 of 6 regions
country_iso3 string ISO 3166 country code

Peril-specific intensity fields (populated based on peril_type):

Hurricane: max_wind_speed_knots, central_pressure_mb, storm_surge_ft, hurricane_category (1-5), rainfall_inches_72hr, inland_penetration_miles

Earthquake: moment_magnitude_mw, peak_ground_acceleration_g, modified_mercalli_intensity, liquefaction_risk, aftershock_sequence_flag, tsunami_trigger_flag

Flood: flood_type, fema_flood_zone, inundation_depth_ft, inundation_area_sq_miles, flood_duration_days, peak_discharge_cfs

Loss decomposition (USD):

Column Description
insured_loss_usd Total insured loss
economic_loss_usd Total economic loss (insured + uninsured)
residential_loss_usd Residential portion
commercial_loss_usd Commercial portion
industrial_loss_usd Industrial portion
auto_loss_usd Auto portion
marine_cargo_loss_usd Marine/cargo portion
business_interruption_loss_usd BI portion
industry_loss_usd Industry-wide loss for trigger purposes

Actuarial pipeline:

Column Description
aal_contribution_usd Average Annual Loss contribution
oep_percentile OEP curve percentile
aep_percentile AEP curve percentile
probable_maximum_loss_pml_pct PML as % of TIV
tail_value_at_risk_tvar TVaR (CVaR)
reinsurance_recoveries_usd Reinsurance recoveries
net_retained_loss_usd Net retained loss (cedant)
cat_bond_trigger_flag yes/no — OEP > 99th percentile
regulatory_stress_test_tier Regulatory stress test classification
solvency_ii_scr_event Solvency II SCR event flag

ep_curve_summary.csv

Column Type Description
peril_type string Peril
return_period_years int Return period (10, 25, 50, 100, 200, 500, 1000)
oep_loss_usd float OEP loss at this return period
exceedance_probability float 1/return_period

Suggested Use Cases

  • Training catastrophe loss prediction models — predict insured loss from intensity features
  • EP curve construction & validation — model OEP/AEP curves at multiple return periods (10-1000 year)
  • Reinsurance pricing models — train layer attachment and recovery models
  • Cat bond trigger prediction — multi-peril 99th-percentile detection
  • Climate scenario stress testing — comparison across baseline / RCP4.5 / RCP8.5 climate scenarios (full product)
  • Peril-specific vulnerability curve fitting by construction type
  • Geographic risk concentration analysis — region × peril modeling
  • Solvency II SCR event classification
  • PML / TVaR computation for portfolio risk
  • Demand surge multiplier modeling post-event
  • Mean damage ratio (MDR) prediction by intensity and construction
  • Wildfire/flood frequency forecasting under climate scenarios
  • Hurricane track-curvature & forward-speed modeling
  • Earthquake liquefaction risk + tsunami trigger correlation
  • Insurtech catastrophe model training without proprietary RMS/AIR licenses

Loading the Data

import pandas as pd

scenarios = pd.read_csv("cat_scenarios.csv")
ep_curve  = pd.read_csv("ep_curve_summary.csv")

# Multi-class peril classification target
y_peril = scenarios["peril_type"]

# Regression: insured loss prediction
y_loss = scenarios["insured_loss_usd"]

# Binary cat bond trigger prediction (rare event ~10-40%)
y_cat_bond = (scenarios["cat_bond_trigger_flag"] == "yes").astype(int)

# Hurricane-only analysis
hurricanes = scenarios[scenarios["peril_type"] == "hurricane"]
hurricane_severity = hurricanes["hurricane_category"]  # 1-5

# Build your own EP curve by peril
peril = "hurricane"
sub = scenarios[scenarios["peril_type"] == peril].sort_values("insured_loss_usd",
                                                              ascending=False)
n = len(sub)
ranks = (n - sub.reset_index().index) / n  # exceedance probability
return_periods = 1 / ranks

License

This sample is released under CC-BY-NC-4.0 (free for non-commercial research and evaluation). The full production dataset is licensed commercially — contact XpertSystems.ai for licensing terms.

Full Product

The full INS-003 dataset includes ~50,000 catastrophe scenarios across all 6 perils, with configurable climate scenarios (baseline / RCP4.5 / RCP8.5), configurable catalog horizons (10,000-100,000 years), and per-peril deep dives for the catastrophe modeling community.

📧 pradeep@xpertsystems.ai 🌐 https://xpertsystems.ai

Citation

@dataset{xpertsystems_ins003_sample_2026,
  title  = {INS-003: Synthetic Catastrophe Scenarios Dataset (Sample)},
  author = {XpertSystems.ai},
  year   = {2026},
  url    = {https://huggingface.co/datasets/xpertsystems/ins003-sample}
}

Generation Details

  • Generator version : 1.0.0
  • Random seed : 42
  • Generated : 2026-05-16 19:59:28 UTC
  • Climate scenario : baseline
  • Catalog horizon : 10,000 years
  • Architecture : Stochastic event catalog with hard actuarial constraints
  • Overall validation: 100.0 / 100 (grade A+)