Customer records available for model development.
Predictive analytics / Customer strategy
Customer Targeting with Predictive Analytics
A classification model is not the deliverable. The deliverable is a credible answer to a business question: which customers should the team approach first, and why?
- Role
- Data Analyst / Junior Data Scientist
- Tools
- Python, Pandas, Scikit-learn, XGBoost
- Challenge
- Highly imbalanced binary target
- Output
- Ranked top 100 customer list
The problem
Prioritize outreach without treating every lead equally.
Relationship managers needed a practical way to identify potential customers most likely to be interested in a mortgage product. Historical data contained useful patterns, but the positive class was rare and simple accuracy would create a misleading sense of performance.
The project had two responsibilities: build a defensible probability model and translate its output into something a customer-facing team could actually use.
Model context
The rare outcome shaped every analytical decision.
A small mortgage class requiring imbalance-aware evaluation.
Potential customers ranked by predicted probability.
Method
Build the model around the decision it must support.
- Understand the customer dataExplored distributions, missing values, categorical variables and potential leakage before modelling.
- Prepare for imbalanceTransformed variables, tested resampling and evaluated models beyond overall accuracy.
- Compare classifiersTrained candidate models and selected the approach that best balanced useful positive detection with stable ranking.
- Score, do not merely labelConverted probabilities into a ranked list that relationship managers could prioritize.
Product thinking
The model output needed an operating loop.
| Model output | Business use | Feedback needed |
|---|---|---|
| Probability score | Prioritize high-propensity leads for outreach. | Actual contact and conversion outcomes. |
| Top customer profile | Shape messaging for the highest-ranked group. | Response differences by segment and offer. |
| Ranked list | Allocate limited relationship-manager time. | Campaign yield by score band. |
Outcome
Predictive work translated into a usable customer strategy.
The final output was a ranked, business-readable lead list with a clear validation plan. The next iteration would connect campaign responses back to the model, compare predicted probability with real conversion and refine targeting by segment.
Next case study: Email Campaign →