End-to-end machine learning pipeline predicting 12-month revenue at signup for a subscription-based insurance product (~91k customers).

Standard regression fails on insurance LTV because revenue is driven by two separate events: whether a customer stays, and whether they buy additional policies. A single model can't capture both cleanly.

Rather than predicting LTV directly, the pipeline decomposes it:

1. Tenure classifier — probability the customer stays 12 months
2. Cross-sell classifier — probability they purchase an additional policy
3. Cross-sell valuator — expected value of that policy if purchased

Formula: LTV = (Monthly Commission × Predicted Tenure) + (P_crosssell × Value_crosssell)

16% reduction in MAE over baseline.

• Customers starting with Product C show the highest cross-sell propensity
• Users on non-standard OS show significantly higher churn — likely UX friction
• Retention drives more total portfolio value than upsell events

Python · Pandas · scikit-learn · XGBoost · Jupyter

pip install -r requirements.txt
jupyter notebook LTV_Model.ipynb

composite-ltv-modeling/
├── LTV_Model.ipynb      # Main modeling notebook
├── requirements.txt
└── README.md