Insurance Fraud Detection and Claim Risk Scoring

About

An evaluation-aware insurance fraud and claim-risk machine-learning pipeline with temporal features, leakage auditing, imbalance handling, threshold tuning, business-loss analysis, serialized artifacts, and a Streamlit scoring interface.

Status: Evaluation-aware local ML pipeline; no production success is claimed.
Stack: Python · pandas · scikit-learn · Streamlit · YAML configuration
Proof: Temporal validation, leakage auditing, imbalance handling, threshold tuning, and business-loss analysis.

This project builds an end-to-end machine learning pipeline for insurance fraud detection and claim severity prediction. It uses time-aware validation, leakage-safe feature engineering, imbalance handling, threshold tuning, business-cost evaluation, and a Streamlit interface for claim risk scoring.

Project Structure

• app/: Streamlit dashboard.
• configs/: YAML configuration for paths, split settings, and model parameters.
• src/data/: Data loading and dataset creation.
• src/features/: Feature engineering logic.
• src/models/: Training, prediction, risk scoring, and anomaly model helpers.
• scripts/: Local validation and model inspection scripts.
• tests/: Automated smoke tests.
• notebooks/Insurance_Fraud_ML_Pipeline_new.ipynb: Main ML notebook.

Local-Only Artifacts

The following folders are intentionally ignored by Git:

• data/: raw/interim/processed datasets.
• models/: trained model binaries and metadata.
• reports/: generated metrics and analysis outputs.
• .venv/: local Python environment.

Do not commit private datasets, model binaries, cache folders, or generated reports.

See dataset provenance notes for the input-data publication boundary.

Setup

Run these commands from the project root:

python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
pip install -r requirements.txt

Place these files in data/raw/:

• insurance_data.csv
• employee_data.csv
• vendor_data.csv

Train Models

python -m src.models.train

Training creates local artifacts under data/processed/, models/, and reports/.

Run Streamlit App

streamlit run app/streamlit_app.py --server.port 8501

Open:

http://localhost:8501

Validate

python scripts/smoke_load.py
python scripts/inspect_model.py
pytest -q

Evaluation snapshot

The documented notebook output at commit 5c5b99564313c9c84021e8e1dce263e1eba99ad6 records a temporal split of 7,000 train rows, 1,500 validation rows, and 1,500 test rows. The selected checkpoint was Autoencoder (Anomaly) with a threshold of 0.20.

Metric	Test-window result
PR-AUC	0.044732
ROC-AUC	0.483760
Recall	0.898551
Precision	0.046757
F2 score	0.193508
Business loss	982,000

These values are a documented notebook checkpoint, not a new run or a claim of production success. The high recall and low precision show the review-capacity trade-off. The notebook's false-negative and false-positive costs are scenario parameters, not verified insurer costs or business savings. See evaluation notes for the interpretation boundary and required run metadata.

Model Selection

The final fraud model is selected using fraud-focused and business-focused metrics, not accuracy. The comparison prioritizes PR-AUC, recall, F2-score, threshold-tuned business loss, and top-K investigation lift. This matches the real use case: ranking suspicious claims for investigation while controlling false-review cost.