You can deploy this model and start earning money today!

sahilali09/
prcp-1010-insurance-claim-prediction

Name: prcp-1010-insurance-claim-prediction
Availability: InStock

ml-intern

Views

No views yet

PRCP-1010: Insurance Claim Prediction

Capstone Project — Predicting which insurance customers will submit a claim to optimize marketing campaigns.

Problem Statement

Task 1: Predictive Model

Build a model to identify customers most likely to submit an insurance claim, enabling the marketing team to prioritize outreach and optimize resource allocation.

Task 2: Marketing Suggestions

Provide actionable recommendations for the insurance marketing team to convert potential customers across different risk segments.

Dataset

Source: Porto Seguro Safe Driver Prediction (Anonymized)
Size: 595,212 observations, 59 anonymized features
Target: Binary (1 = claim submitted, 0 = no claim)
Class Distribution: 96.4% no claim, 3.6% claim (highly imbalanced)
Missing Values: Encoded as -1 across multiple features

Models Evaluated

Rank	Model	ROC-AUC	F1-Score	Accuracy
1	Gradient Boosting	0.6276	0.0009	0.9635
2	Random Forest	0.6222	0.0975	0.6818
3	Logistic Regression	0.6162	0.0000	0.9635
4	LightGBM	0.6102	0.0730	0.1278
5	Naive Bayes	0.6093	0.0962	0.8588
6	XGBoost	0.6049	0.0737	0.1436
7	Decision Tree	0.5810	0.0000	0.9636
8	K-Nearest Neighbors	0.5336	0.0233	0.9549

Best Model: Gradient Boosting

Primary Metric: ROC-AUC = 0.6276
Handles mixed data types natively (categorical, binary, continuous)
Robust to outliers and missing values through tree-based splits
Interpretable feature importance for marketing insights
Fast inference suitable for production customer scoring

Key Challenges & Solutions

Challenge	Technique	Reason
Missing values encoded as `-1`	Replace with NaN, median/mode imputation	Standardizes missing handling; prevents sentinel distortion
Severe class imbalance (96.4/3.6)	Class weights, `scale_pos_weight=26`, ROC-AUC metric	Prevents majority-class bias; measures ranking quality
Anonymized features	Group-based feature engineering (`ps_ind_sum`, `ps_car_sum`, `missing_count`)	Extracts structural signal without semantic meaning
Large dataset (595K rows)	Stratified subsampling for training; full test for evaluation	Balances speed vs. reliable metrics
High cardinality categoricals	Tree-based models (native categorical handling)	Avoids one-hot dimensionality explosion
Noisy calculated features	Monitor feature importance empirically	Tree models ignore irrelevant features automatically
Metric selection	ROC-AUC as primary (ranking), F1/Accuracy as secondary	Aligns with marketing prioritization objective

Marketing Recommendations

High-Risk Segments (Top 3 Deciles)

Proactive personalized outreach before competitor research
Premium products, flexible payments, dedicated advisors
Emphasize peace of mind and fast claim processing

Medium-Risk Segments (Deciles 4-7)

Education-first campaigns with testimonials and case studies
Free risk assessments as lead magnets
Retargeting ads and chatbot-assisted quotes

Low-Risk Segments (Bottom 3 Deciles)

Cost-leadership with basic plans and bundle discounts
Referral incentive programs
Self-service digital channels

Files

PRCP_1010_Insurance_Claim_Prediction.ipynb — Main Jupyter notebook with all tasks
best_insurance_claim_model.pkl — Trained production model
feature_scaler.pkl — StandardScaler artifact
feature_names.pkl — Expected feature order
Various .png — Analysis and visualization plots

Usage

python

import joblib
import pandas as pd

model = joblib.load('best_insurance_claim_model.pkl')
features = joblib.load('feature_names.pkl')

# Predict claim probability for new customers
proba = model.predict_proba(new_data[features])[:, 1]

Project: PRCP-1010 | Domain: Finance/Insurance

Generated by ML Intern

This model repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

Try ML Intern: https://smolagents-ml-intern.hf.space
Source code: https://github.com/huggingface/ml-intern

sahilali09/prcp-1010-insurance-claim-prediction