Travelers Insurance Fraud Detection Pipeline (NESS 2025)

This project was developed for the 2025 New England Statistics Symposium (NESS) Statathon, where participants were tasked with building a predictive model to identify first-party physical damage fraudulence in auto insurance claims. Our solution placed 2nd overall.

The codebase is structured to support general-purpose binary classification tasks, with modular components for training, evaluation, and ensemble modeling. Though initially developed for the Travelers Insurance fraud dataset, the pipeline is designed to be reusable with other structured/tabular classification problems.

---

Problem Statement

As a modeling team at Travelers Insurance, the goal was to:

1. Predict fraudulent claims related to first-party physical damage using historical data
2. Identify and interpret key drivers of fraud
3. Provide actionable insights for improving claim handling procedures

---

Input

• Tabular data in .csv format, consisting of structured features (numerical, categorical)
• No unstructured data (text/images) included in this version

---

Models Used

• Gradient Boosting Models
  • XGBoost, LightGBM, CatBoost, HistGradientBoostingClassifier
• Linear Baseline
  • Lasso (via LogisticRegression with L1 penalty)
• Model Ensembling
  • Weighted averaging of calibrated model outputs

---

Evaluation

• Metric: F1 Score (targeting balance between precision and recall for fraud)
• Validation Strategy: K-Fold Cross-Validation (5-fold by default)