ic_008 · Industry: Insurance — claim fraud detection
Status: ✓ Passing · Last run: 2026-06-07T07:35:28 · Pipeline: v2
· History: 1/2 runs passed (50%)
What was tested
Insurance — claim fraud detection Fraud detection: policies + claims + claim_lines + customer_history + labels. Signal: frequent claims + high charges.
What we planted in the data
- 300 entities
- Signal strength: 0.70 (sigmoid slope multiplier)
What the system did
- Training rows built: 1,200
- Features used in model: 24
- Model selected:
RandomForestClassifier - CV AUC: 0.732
Top features by importance:
claim_count_365d(0.117)claim_amount_sum_365d(0.112)claim_amount_agg_avg_by_customer(0.106)claim_amount_max_180d(0.104)claim_count_180d(0.080)
What we checked — all assertions passed
| Status | Assertion | Detail |
|---|---|---|
| ✓ | no_pipeline_errors |
clean run |
| ✓ | feature_survival_rate_above |
24/60 features survived (40% ≥ 30%) |
| ✓ | source_tables_contributing_above |
6 distinct source-table prefixes in surviving features (['claim', 'history', 'policy', 'premium', 'prior', 'tenure']) |
| ✓ | auc_in_band |
AUC = 0.732 inside band [0.55, 0.95] |
How this could have gone wrong
(no assertion descriptions provided)
Why this case matters
Industry-canonical ML problem; multi-table shape mirrors real customer datasets.
Reproducing
# from auto_insight_api/
python -m validation.v2 run ic_008 --pipeline v2 -v
- Case config:
validation/v2/cases/ic_008_insurance_fraud.yaml - Data shape:
industry_insurance_fraud - Analytics type:
ml_binary