Neural network models on transaction data of a settlement account

Dmitry Koptelov4 Jun 2022

Problem

A bank wants to pick customers who can get a loan with the lowest default risk, where default means 90 days past due or a restructured loan within a year. A tabular boosting model already exists; the question is whether account transactions add signal. Unlike a card transaction (terminal, time, merchant category code (MCC), region), a settlement account transaction has only a date and an amount. The payment purpose sits in a text description mixing bank templates with free user text, so rules cannot categorize it.

Example from the article: Alfa-Bank credit scoring on settlement account transactions, where the neural model blended with the boosting model raised Gini by about 2 points on secondary sales.

Solution in brief

Clean the description by masking numbers and names, lemmatizing and keeping only popular tokens. Train Word2Vec on the descriptions and feed each transaction embedding into a bidirectional recurrent network with min/max pooling. Blend its score with the boosting score using logistic regression, and report the gain over boosting alone.

Questions this case answers
  • How to turn noisy free-text fields of event records into features for a sequence model when no rules can cover the text generation process
  • How to combine a neural sequence model with an existing tabular boosting model and measure the incremental quality of the blend
  • How to use free-text payment descriptions from account transactions as features in a sequence model?
  • How to build a credit scoring model from settlement account transactions that carry only date, amount, direction and a text description?

Tools and methods

  • Credit scoring — default probability within a year, reported as Gini.
  • Word2Vec — 50-dimensional token embeddings trained on the corpus of cleaned payment descriptions.
  • Topic modeling — checks that descriptions carry information.
  • Bidirectional RNN with min/max pooling — reads the transaction sequence of a customer.
  • Blending with logistic regression — weights the boosting score and the neural network score.
Continue reading — it's free

Google or a one-time email code. No password.