README.md

metadata

license: cc-by-nc-4.0
task_categories:
  - tabular-classification
  - tabular-regression
tags:
  - insurance
  - fraud-detection
  - claims
  - siu
  - anomaly-detection
  - graph-ml
  - social-network-analysis
  - synthetic-data
  - p-and-c
  - red-flag-scoring
pretty_name: INS-007 — Synthetic Insurance Fraud Claims Dataset (Sample)
size_categories:
  - 1K<n<10K

INS-007 — Synthetic Insurance Fraud Claims Dataset (Sample)

XpertSystems.ai Synthetic Data Platform · SKU: INS007-SAMPLE · Version 1.0.0

This is a free preview of the full INS-007 — Synthetic Insurance Fraud Claims Dataset product. It contains roughly ~10% of the full dataset at identical schema, fraud taxonomy, and SIU investigation workflow, so you can evaluate fit before licensing the full product.

File	Rows (sample)	Rows (full)	Description
fraud_claims.csv	~5,000	~50,000	Per-claim records with fraud labels (99 cols)
fraud_ring_edges.csv	~435	~5,000	Network edge list for graph-based fraud ML

Dataset Summary

INS-007 simulates the full fraud claims investigation lifecycle — from claim intake through SIU referral, surveillance, forensic investigation, and adjudication — with realistic ring-based fraud structures, staged accident typologies, and ISO ClaimSearch-calibrated red flag scoring.

Calibration sources:

• Coalition Against Insurance Fraud — confirmed fraud rates, savings
• FBI Insurance Fraud Annual Report — organized ring fraud share
• NAIC AS-101 — Anti-Fraud Plan regulations and SIU referral rates
• Insurance Research Council (IRC) — attorney representation rates
• ISO ClaimSearch — red flag indicator catalog and scoring
• NICB ForeWARN — staged accident pattern typologies

9 lines of business with empirically-anchored fraud rates:

• Auto bodily injury (28%) — highest fraud rate LOB
• Auto property damage (18%)
• Auto PIP / no-fault (14%) — high fraud rate in no-fault states
• Workers compensation (12%)
• Homeowners (10%)
• Commercial property (6%)
• Health (5%)
• General liability (4%)
• Disability (3%)

8 staged accident pattern typologies (NICB-calibrated):

• Swoop-and-squat
• Drive-down
• Side-swipe
• Rear-end setup
• Paper accident (no actual collision)
• Owner give-up (insurance dump)
• Medical mill only (no accident at all)

12 fraud type confirmations (post-investigation):

• Staged accident
• Soft fraud / exaggeration
• Medical billing fraud
• Provider fraud (corrupt medical / repair providers)
• Agent fraud (corrupt agent / broker)
• Owner give-up
• Workers comp fraud
• Disability fraud
• Slip and fall fraud
• Fire loss fraud
• Water damage fraud
• Ring-organized fraud (multi-party coordinated)

Fraud ring social network:

• 60 fraud rings in sample (full product: ~150 rings)
• Average ring size 5-15 members
• Shared attributes: addresses, phone numbers, attorneys, body shops, medical providers, witnesses
• 435 network edges in sample (full product: ~5,000)
• Edge types: same_ring, shared_address, shared_phone, shared_attorney, shared_body_shop, shared_medical_provider, shared_witness

SIU investigation workflow:

• SIU referral with reason taxonomy
• Surveillance conducted flag
• Recorded statement taken
• Statement inconsistency count
• Investigation duration
• Outcome (paid_no_siu, paid_post_siu_legitimate, settled_with_reduction, denied_unsubstantiated, denied_fraud_confirmed, withdrawn_by_claimant, referred_to_da, criminal_prosecution, civil_recovery)
• Savings achieved (avoided loss)
• Recovery amount (post-payment clawback)

Red flag scoring (40+ binary indicators with prefix rf_):

• Claimant red flags: prior_claims_history, gaps_in_treatment, no_witnesses
• Provider red flags: medical_mill_pattern, billing_anomalies, shared_attorney_concentration
• Vehicle red flags: salvage_history, pre_existing_damage, phantom_vehicle
• Network red flags: shared_address, shared_phone, shared_body_shop
• Statement red flags: inconsistent_statements, recorded_changes
• Behavioral red flags: late_reporting, policy_lapse_pre_loss, coverage_recently_increased
• ISO ClaimSearch-calibrated weights

Red flag composite score (0-100):

• Computed weighted score across all triggered red flags
• Cleanly separates fraud (mean 50.1) from legitimate (mean 7.2)
• Direct feature for ML model training

Claim-level features:

• Claim ID with structured format (CLM-YEAR-LOB-STATE-SEQ)
• Loss date, claim filing date, age of claim
• Claim amount demanded, paid, savings, recovery
• Medical billing amount
• Estimated fraud savings (counterfactual avoided loss)

Calibrated Validation Results

Sample validation results across 10 fraud-detection KPIs:

Metric	Observed	Target	Source	Verdict
n_states_represented	50	12	Min state coverage (national mix)	✓ PASS
n_lob_represented	9	9	9 lines of business	✓ PASS
confirmed_fraud_rate_pct	10.14	10.00	Coalition Against Ins Fraud	✓ PASS
siu_referral_rate_pct	16.12	16.00	NAIC AS-101 SIU referral rate	✓ PASS
ring_based_fraud_share_pct	9.90	10.00	FBI organized fraud share	✓ PASS
attorney_representation_pct	30.10	30.00	IRC attorney rep (high-suspicion)	✓ PASS
surveillance_conducted_pct	10.70	10.00	SIU surveillance industry rate	✓ PASS
police_report_filed_pct	71.84	70.00	Auto-heavy LOB police report rate	✓ PASS
red_flag_separation_ratio	6.938	6.000	RF score: confirmed/legit ratio	✓ PASS
n_fraud_ring_edges	435	200	Network edge density (floor)	✓ PASS

Note: Confirmed fraud rate (~10%) is below the configured --fraud-rate 0.18 parameter because only ~56% of fraud-suspect claims that get SIU referral end in fraud_confirmed=True after investigation. The other ~44% of suspected fraud is resolved as paid_post_siu_legitimate, settled_with_reduction, denied_unsubstantiated, or withdrawn — realistic SIU workflow outcomes. This makes INS-007 useful for training models that distinguish suspected from confirmed fraud, not just legitimate from suspicious.

Schema Highlights

fraud_claims.csv (primary file, 99 columns)

Claim identification:

Column	Type	Description
claim_id	string	Structured ID: CLM-YEAR-LOB-STATE-SEQ
policy_id, claimant_id	string	Foreign keys
incident_state	string	U.S. state code
loss_date, claim_filed_date	date	Timeline
lob	string	1 of 9 LOBs
incident_type	string	10 incident types
staged_accident_pattern	string	8 NICB typologies (or 'none')

Claim financials:

Column	Type	Description
claim_amount_demanded_usd	float	Initial demand
claim_amount_paid_usd	float	Amount paid (post-investigation)
medical_billing_usd	float	Medical billing component
savings_achieved_usd	float	Avoided loss from SIU intervention
recovery_amount_usd	float	Post-payment recovery
estimated_fraud_savings_usd	float	Counterfactual full-fraud cost

Investigation workflow:

Column	Type	Description
siu_referral_flag	bool	Referred to SIU
siu_referral_reason	string	Categorical reason
surveillance_conducted	bool	SIU conducted surveillance
recorded_statement_taken	bool	Recorded statement taken
statement_inconsistency_count	int	Inconsistencies detected
investigation_duration_days	int	Days SIU investigated
police_report_filed	bool	Police report on file
attorney_represented	bool	Claimant attorney representation
attorney_siu_flag	bool	Attorney on SIU watchlist
claim_outcome	string	9-class outcome

Fraud labels (training targets):

Column	Type	Description
fraud_confirmed	bool	Binary fraud label (post-investigation)
fraud_type_confirmed	string	12-class fraud type (or 'none')
fraud_ring_id	string	Ring ID (nullable)
referral_to_law_enforcement	bool	Referred to law enforcement
criminal_charges_filed	bool	Charges filed
civil_recovery_initiated	bool	Civil recovery action

Red flag indicators (40+ binary columns, prefix rf_):

Examples: rf_prior_claims_history, rf_late_reporting, rf_inconsistent_statements, rf_medical_mill_pattern, rf_no_witnesses, rf_policy_lapse_pre_loss, rf_coverage_recently_increased_flag, rf_phantom_vehicle, rf_shared_attorney, rf_shared_body_shop, rf_treatment_gap, rf_billing_anomalies, rf_shared_address, rf_shared_phone_number, rf_injury_vs_damage_paradox.

Composite scoring:

Column	Type	Description
red_flag_score	float	Composite 0-100 (ISO-calibrated weights)
red_flag_count	int	Number of triggered red flags
fraud_probability_score	float	Internal Bayesian posterior (training tgt)

fraud_ring_edges.csv (network graph)

Column	Type	Description
ring_id	string	Fraud ring identifier (RING-001 to RING-060)
claim_id_a	string	Edge source claim
claim_id_b	string	Edge target claim
edge_type	string	same_ring / shared_address / shared_phone / ...

Suggested Use Cases

• Binary fraud classification — predict fraud_confirmed from claim features
• Multi-class fraud typing — predict fraud_type_confirmed (12 classes)
• SIU referral prediction — predict which claims need investigation
• Red flag score modeling — train regression on red_flag_score
• Staged accident pattern detection — multi-class typology classification
• Fraud ring detection — graph-based community detection
• Network anomaly detection — find suspicious link patterns
• Provider fraud detection — medical mills, body shops, attorneys
• Soft fraud vs hard fraud classification
• Attorney representation prediction for high-suspicion claims
• Claim adjudication automation — predict 9-class claim_outcome
• Savings optimization — maximize avoided loss per SIU investigation hour
• Fraud ring growth modeling — track ring expansion over time
• Surveillance prioritization — predict which claims warrant surveillance
• Statement inconsistency detection — NLP on recorded statements
• Insurtech SIU model training without proprietary ISO ClaimSearch data
• Graph neural network (GNN) training with realistic fraud ring topology

Loading the Data

import pandas as pd

claims = pd.read_csv("fraud_claims.csv")
edges  = pd.read_csv("fraud_ring_edges.csv")

# Binary fraud classification target
y_fraud = claims["fraud_confirmed"].astype(int)

# Multi-class fraud type (12 classes; filter to fraud only)
fraud_only = claims[claims["fraud_confirmed"] == True]
y_fraud_type = fraud_only["fraud_type_confirmed"]

# Binary SIU referral prediction
y_siu = claims["siu_referral_flag"]

# Multi-class staged accident pattern (8 classes)
y_staged = claims["staged_accident_pattern"]

# Regression: red flag score
y_rf_score = claims["red_flag_score"]

# 9-class claim outcome
y_outcome = claims["claim_outcome"]

# Build fraud ring graph for GNN training
import networkx as nx
G = nx.Graph()
for _, row in edges.iterrows():
    G.add_edge(row["claim_id_a"], row["claim_id_b"],
               edge_type=row["edge_type"], ring_id=row["ring_id"])
print(f"Ring graph: {G.number_of_nodes()} nodes, {G.number_of_edges()} edges")

# Compute community detection features
import community as community_louvain  # pip install python-louvain
partition = community_louvain.best_partition(G)

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-007 dataset includes ~50,000 fraud claims records across 99 columns, with ~5,000 network edges spanning ~150 fraud rings. Calibrated to Coalition Against Insurance Fraud, FBI fraud reports, NAIC AS-101, IRC, and ISO ClaimSearch.

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

Citation

@dataset{xpertsystems_ins007_sample_2026,
  title  = {INS-007: Synthetic Insurance Fraud Claims Dataset (Sample)},
  author = {XpertSystems.ai},
  year   = {2026},
  url    = {https://huggingface.co/datasets/xpertsystems/ins007-sample}
}

Generation Details

• Generator version : 1.0.0
• Random seed : 42
• Generated : 2026-05-16 20:39:10 UTC
• Records : 5,000 (~10.1% confirmed fraud)
• Rings : 60 fraud ring structures
• Calibration basis : Coalition Against Ins Fraud + FBI + NAIC AS-101 + IRC + ISO ClaimSearch
• Overall validation: 100.0 / 100 (grade A+)