Problem
A bank wants to lend to the customers with the lowest default risk. The source data is a time series of customer events: card transactions, purchases, communications and mobile app logs. Such sequences can be arbitrarily long, so classical models need a data scientist to invent many aggregates, which takes time and loses information. A second problem is the label: default is observed for a year after the loan, so the newest loans cannot be used for training.
Example from the article: Alfa-Bank's credit scoring models predict the probability that a loan goes 90 days past due, or is restructured, within a year of issue.
Solution in brief
Feed the raw event sequence to a neural network. Each event becomes an embedding: numerical features are binned by quantiles into categories, and every categorical feature gets a trainable embedding that is concatenated with the others. To reduce the label delay, the model is validated out of time and on shorter default definitions that correlate with the classical one.
Questions this case answers
- How to represent a long sequence of mixed categorical and numerical events as model input without manual feature engineering
- How to train and validate a risk model when the outcome label matures only after a long delay
- How to turn a customer's raw sequence of bank events into model input without hand-built aggregates?
- How to embed numerical and categorical event features in one sequence model?
Tools and methods
- Entity embeddings — trainable embedding layer per categorical feature, learned end-to-end with the task.
- Quantile binning — turns numerical features into categorical ones.
- Out-of-time validation — holds out the latest loans to check the model on "future" data.
- Shortened default definitions — 60 days within 6 months and similar, to cut the label delay.
Google or a one-time email code. No password.