The question

How can campaign data support more informed marketing decisions? I combined two supplied datasets into 33,908 customer records, checked missing values and examined the distributions of age, balance and call duration.

The approach

I used a 70/30 training and testing split to compare logistic regression with a CHAID decision tree. The analysis considered customer characteristics and contact information, then compared how well each model ranked campaign responses.

What I found

Both models achieved a reported test AUC of 0.860. I preferred the tree as a communication tool because its visual groups made the patterns easier to explain. This is a result on the assignment dataset, not a measured improvement in a live campaign.

ROC curves comparing logistic regression and a CHAID decision tree in my assignment.

An important limitation

Call duration was a strong predictor, but it is only known after contact. These models therefore cannot be treated as ready-to-use pre-call targeting models. Their association between longer calls and response also does not establish that deliberately extending a call will cause a conversion.

The project demonstrates model comparison and interpretation. A practical next step would be to evaluate a model using only information available before contact.

Skills explored

  • SPSS
  • Data preparation
  • Logistic regression
  • CHAID decision trees
  • Model evaluation