The-House-Of-Kgosi/Short-Term-Insurance-Churn-in-South-Africa

# 🇿🇦 SA Short-Term Insurance — Customer Churn Prediction ### Machine Learning · South African Market Context · Vehicle & Home Insurance --- ## Overview Customer churn — when a policyholder cancels or fails to renew their policy — is one of the most costly problems facing South African short-term insurers. Research consistently shows that retaining an existing customer costs **5–7× less** than acquiring a new one. This project builds and compares three machine learning classification models to predict churn probability for SA vehicle and home insurance customers, using features grounded in the local market reality: **load-shedding exposure, premium-to-income pressure, provincial risk profiles, claims disputes, and broker vs. direct channel dynamics.** The model produces actionable churn risk scores that an insurer can use monthly to proactively target at-risk customers before they cancel. --- ## South African Market Context This project deliberately reflects features unique to the SA insurance landscape: | SA-Specific Feature | Why It Matters | |---|---| | **Load-shedding stage** | Stages 4–6 correlate with surge/fire claims, frustration, and price-shopping | | **Premium-to-income ratio** | SA's cost-of-living crisis makes this the #1 churn driver | | **Provincial risk tier** | Gauteng, KZN and Western Cape carry distinct risk and churn profiles | | **Disputed claims flag** | SA policyholders with unresolved disputes churn at nearly 2× the average rate | | **Broker vs direct channel** | Broker-acquired customers churn significantly less due to relationship management | | **Competitor quotes obtained** | A leading indicator of imminent cancellation — key early warning signal | --- ## Models Compared | Model | AUC Score | Notes | |---|---|---| | **Logistic Regression** | Best performer | Interpretable, good for regulatory explainability requirements | | **Random Forest** | Strong | Handles non-linearity, useful feature importance output | | **Gradient …