0.809 to 0.829Real bank data from an open research dataset
Signals that earn their place
AUC for product take up across 100,224 business clients, with account history signals added
- What was observed
- Business clients' transactions and the products they took up the next month.
- What a generic approach says
- A strong model on summary features: AUC 0.809.
- What the engine read
- Adding signals built from each client's own history over time raised it to 0.829, with a spread of 0.008 across folds.
- What it meant for the decision
- Signals built from how an entity behaves over time add to a strong model rather than repeat it. The outcome here is product take up, not default, and the test was across clients, not out of time.