Back to projects

Machine Learning & Analytics

Multi-Output Insurance Purchase Prediction Pipeline

Predicting all 7 insurance policy options at once, measured against an honest baseline.

model, average per option
92.54%model, average per option
last-quote baseline
92.48%last-quote baseline
options beat the baseline
6 / 7options beat the baseline
test customers
19,402test customers

// at a glance

The problem and the idea

The problem

In an insurance quote funnel, knowing which options a customer will finally buy helps with pricing and cross-selling. On this data, simply predicting the last quote the customer viewed is already about 92.5% accurate, so a model's headline accuracy can look impressive while adding almost nothing.

The idea

Treat the 7 policy options as one multi-output classification problem, engineer features from each customer's quote journey, and judge every model against the last-quote baseline. The real question is whether a model can learn when a customer will move away from their last quote.

// who it's for

Who it's for

  • 01

    Insurance analytics and pricing teams

  • 02

    Recommendation systems that need to know when a customer will change their mind

  • 03

    ML practitioners interested in honest evaluation

// how it works

How it works

quote logsjourney featuresMultiOutput XGBoptions A–Glast-quote baseline92.54% vs 92.48% baseline · biggest gain on target G

1/5 Exploratory analysis

Swipe sideways to see the whole diagram →

// key features

Key features

  • Seven targets (A–G) predicted by one model
  • A baseline comparison for every option, not accuracy in a vacuum
  • Per-target overfitting check: train–test gap of 1 point or less
  • Deep dive on the hardest target with SHAP explanations
  • Inference demo that decodes the full 7-option package

// design decisions

Design decisions

  1. 01

    The baseline is reported next to every result, even though it makes the gain look small.

  2. 02

    The model's value is concentrated where it matters: the hardest target and customers with unstable histories.

  3. 03

    Fitted preprocessing lives in the modeling notebook, so nothing leaks from the test set.

// built with

Built with

  • Python
  • XGBoost
  • CatBoost
  • scikit-learn
  • SHAP
  • Feature Engineering

Run it locally

$ pip install -r requirements.txt