← Back to selected work
Machine Learning · Case Study

Travel Insurance Purchase Prediction

End-to-end machine learning

A leakage-conscious classification project focused on model comparison, honest evaluation, interpretation and real-world decision limits.

PythonPandasscikit-learnClassificationCross-validationModel evaluation
Travel Insurance Purchase Prediction project cover
Download project archiveGitHub profile
01 / Question

What is the real decision?

Can historical customer characteristics help identify likely travel-insurance purchasers while keeping the performance estimate honest enough to support a targeting decision?

Framing

The project separates exploration and inference from predictive modeling. A stratified holdout is reserved before model selection.

Why it matters

A model is only useful if the estimated performance survives contact with genuinely unseen data and if the business cost of errors is understood.

02 / Method

Evaluation before complexity

Workflow

  • Preprocessing inside scikit-learn pipelines
  • Stratified cross-validation
  • Logistic Regression, RBF SVM and Random Forest comparison
  • Training-only hyperparameter tuning
  • Final evaluation on an untouched holdout

Evaluation

  • ROC-AUC and PR-AUC
  • Precision, recall and F1
  • Balanced accuracy
  • Confusion matrix and error analysis
  • Permutation importance after evaluation
03 / Discipline

Protect the test set

The holdout is not used for scaling, encoding, model choice, hyperparameter selection or threshold selection. The purpose is to preserve a genuine final test rather than quietly turning the test set into another development set.
04 / Limits

Where the model can fail

Data limits

  • Historical patterns may not transfer to another market or year.
  • The dataset omits pricing, destination, trip duration, claims and marketing-touchpoint variables.
  • Current-population validation would be required before operational use.

Decision limits

  • A 0.50 threshold is conventional, not economically optimized.
  • False positives and false negatives have different business costs.
  • Age and health-related features would require fairness/compliance review in real insurance use.
Practical interpretation: use the model as a ranking or targeting aid only after validating it on the current customer population and defining the economics of errors. Predictive performance alone is not a deployment decision.