Architecture lab / 06
Customer Churn PredictionUnder the hood.
A train-only preprocessing pipeline, validation-selected threshold, and untouched held-out evaluation keep the experiment boundaries explicit.
Reproducible ML experimentSelect a component to inspect
Control boundary
Stratified split
60% train · 20% valid · 20% test
Separates training, validation, and held-out test rows using recorded seeds.
- Receives
- Features and labels.
- Produces
- Three disjoint partitions.
Design boundary
Test rows are reserved for final reporting, not threshold selection.
Connections
BankChurners dataset Stratified splitlabel & divide
Stratified split Training-only transformstrain partition
AI Services & code Data & output Controls- - Verification path
The core decision
Fit on training data. Choose on validation. Report on test.
Scope & limits
One random split of a public dataset, with no temporal or external validation. Probabilities are not calibrated real-world churn rates. There is no deployed model or creditworthiness inference.
Churn experiment & recorded results