Car Insurance Claim Prediction

This project explores two car-insurance datasets and builds models to predict whether a policyholder will file a claim. The work is organized in two Jupyter notebooks:

• Portfolio_Project.ipynb — Vehicle/specs dataset (highly imbalanced target)
• Car_Claims_Analysis.ipynb — Driver-profile dataset (balanced target)

Datasets

1. Vehicle/specs dataset (imbalanced)

• Features: vehicle specs, equipment flags, regions, etc.
• Target: claim_status (≈6.4% positives)

2. Driver-profile dataset (balanced)

• Features: demographics, driving history, credit and mileage
• Target: OUTCOME (~31% positives)

Approach

• Keep EDA visuals on raw data.
• Split train/test before any target-aware transforms.
• Encoding
  • High-cardinality location → target encoding with cross‑fitting
    • region_code in the first dataset
    • POSTAL_CODE in the second dataset
  • Remaining categoricals → One‑Hot Encoding.
• Missing values
  • In pipelines only (not pre-filled in data)
  • Numeric: median by default; for bell‑shaped variables like CREDIT_SCORE we use mean imputation
  • Categorical: most_frequent
• Threshold selection via the Precision–Recall curve

Models and key results

1. Vehicle/specs (Portfolio_Project.ipynb)

• Baselines
  • Logistic Regression (class_weight='balanced'): ROC‑AUC ≈ 0.62, PR‑AUC ≈ 0.09
  • Random Forest (tuned, class_weight='balanced'): ROC‑AUC ≈ 0.66, PR‑AUC ≈ 0.11
• Notes
  • Strong class imbalance makes precision low at useful recall.
  • After correlation analysis, near‑deterministic equipment flags were dropped; region_code was target‑encoded.

2. Driver profile (Car_Claims_Analysis.ipynb)

• Logistic Regression (balanced data): ROC‑AUC 0.906, PR‑AUC 0.812, Accuracy 0.842
  • Confusion (0.5): TN=1224, FP=149, FN=168, TP=459
• Random Forest: ROC‑AUC 0.897, PR‑AUC 0.787, Accuracy 0.829
  • Confusion (0.5): TN=1193, FP=180, FN=162, TP=465
• Takeaway
  • LR has better ranking (PR/ROC‑AUC), higher precision, fewer false positives.
  • RF finds slightly more true claims (465 vs 459) with more false alarms.

Find the project poster and similarly descriptive materials for this project here!