bdr-ai-org/insurance-motor-claims-decision-v1

Motor Insurance Claims Decision Support Dataset v1

Dataset Description

This is a synthetic dataset designed for training decision support models in motor insurance claims processing. It contains 800 records of motor insurance claims with associated decision labels (approve/review/reject).

Purpose: Human-in-the-loop claims decision support system training.

Schema

The dataset contains the following fields:

Field	Type	Description
claim_id	string	Unique claim identifier (format: CLM-2025-XXXXXX)
claim_amount	float	Amount claimed by policyholder (USD)
vehicle_age	int	Age of vehicle in years (0-24)
accident_type	category	Type of incident: collision, theft, vandalism, weather_damage, fire
police_report	yes/no	Whether a police report was filed
repair_estimate	float	Professional repair cost estimate (USD)
prior_claims	int	Number of prior claims by policyholder (0-5)
decision	category	Recommended decision: approve, review, reject
decision_reason	text	Plain language explanation for the decision

Data Statistics

• Total Records: 800
• Decision Distribution:
  • Review: 584 (73%)
  • Approve: 178 (22.25%)
  • Reject: 38 (4.75%)
• Accident Types: Evenly distributed across 5 categories
• Police Reports: Yes: 534 (66.75%), No: 266 (33.25%)

Data Generation Logic

The dataset is fully synthetic and generated using rule-based logic to ensure consistency:

Decision Rules Applied:

REJECT criteria:

• High claim amount (>$50,000) without police report
• Excessive prior claims (≥4)
• Claim amount significantly exceeds repair estimate (>150%)

APPROVE criteria:

• Low amount (<$5,000) with police report, clean history, and matching estimates
• Standard claims with proper documentation and ≤1 prior claim

REVIEW criteria (default for most cases):

• High value claims (>$30,000)
• Theft or fire incidents (require investigation)
• Old vehicles (>15 years) with high claims
• Multiple prior claims (≥2)

Correlations Built In:

• Newer vehicles tend to have fewer prior claims
• Serious accidents (theft, fire, collision) more likely to have police reports
• Repair estimates vary by accident type and vehicle age
• Claim amounts typically close to repair estimates with normal variation

Limitations

1. Synthetic Data: Not based on real insurance claims; patterns may not reflect actual claim distributions
2. Simplified Rules: Real-world claims involve many more factors (policy details, fraud indicators, medical reports, etc.)
3. No Temporal Data: Missing claim dates, processing times, seasonal patterns
4. Limited Geography: No location-based risk factors
5. No Policy Information: Missing coverage limits, deductibles, policy type
6. Balanced Accident Types: Real-world distributions would be skewed
7. No Fraud Indicators: Missing suspicious pattern detection
8. English Only: Decision reasons in English only

Intended Use

✅ Appropriate Uses:

• Training explainable ML models for claims triage
• Prototyping decision support interfaces
• Testing model interpretability techniques
• Educational purposes for insurance AI systems
• Benchmarking classification algorithms

❌ Inappropriate Uses:

• Production claims processing without human review
• Actual insurance underwriting decisions
• Training models for real-world deployment without domain expert validation
• Any use that could impact real policyholders

Ethical Considerations

• This dataset is for decision support only, not autonomous decision-making
• Human review is mandatory for all claims
• Models trained on this data should not be used to deny claims without human oversight
• The synthetic nature means it lacks real-world bias patterns but also lacks real-world complexity
• No personal or sensitive information is included

Citation

If you use this dataset, please cite:

Motor Insurance Claims Decision Support Dataset v1 (2026)
Generated for AI decision support system development
DEEVO AI Organization

License

MIT License - Free to use for research and development purposes.

Disclaimer

This dataset is for decision support and training purposes only. All decisions must be reviewed and approved by qualified insurance professionals. This data does not represent real claims and should not be used as the sole basis for any insurance decisions.