Public benchmark · reproducible workflow
Bank Marketing
Decision Analytics
A clear path from messy campaign data to a measured review-prioritization model.
This project uses public benchmark data. It does not describe Huntington customers or performance.
Decision question
Can a pre-contact model rank records for review while keeping leakage, uncertainty, and model limits visible?
01 · Data explorer
Start with the data boundary.
The archive has a labeled training file and an unlabeled test file. The reported metrics use a labeled 70/30 split. The archive test file is not used for metric evaluation.
Feature contract
Pre-contact information only
- Kept: customer profile, contact context, campaign history, and economic context.
- Transformed:
pdays=999becomes an explicit prior-contact state. - Excluded: post-call
duration, audit identifiers, and the target label.
Feature experiments
Small changes earn their place.
| Design | ROC-AUC | AP |
|---|---|---|
| Loading… | ||
These comparisons use an internal validation partition. The final test remains reserved for the approved model.
02 · Model and validation
Measure ranking and probability quality.
Logistic regression is the transparent reference. One boosted tree tests whether nonlinear patterns add useful final-test performance.
| Model | ROC-AUC | AP | Brier | Log loss |
|---|---|---|---|---|
| Loading… | ||||
Higher ROC-AUC and AP are better. Lower Brier score and log loss are better.
Benchmark decision
Keep both views.
The boosted tree improves average precision and probability error on this final test. Logistic regression remains the easier model to explain and audit.
This is model behavior on public data. It is not a causal or production claim.
Review capacity example
Use scores to prioritize review, not to automate decisions.
| Loading… |
03 · Methodology and action
Turn model output into a testable next step.
The correct output is not always a stronger model. It may be better data capture, a controlled pilot, or a clear review rule.
Instrument
Record when a campaign requirement became known. Keep the feature-availability boundary visible.
Test
Use a controlled pilot to test whether a review step changes outcomes. A model association is not an intervention effect.
Monitor
Check calibration, drift, segment behavior, overrides, and workload before considering stronger automation.
Required caveat
What this project can and cannot say
It demonstrates cleaning, exploratory analysis, feature engineering, model comparison, uncertainty, and communication. It cannot establish causes of bank outcomes. The source is a historic public retail-bank marketing benchmark, not enterprise-payment data.