3
$\begingroup$

I would like to create a behavioral credit scoring model to score the applications for which transaction data is available. There's an obvious problem mentioned in Thomas et al. Credit Scoring and Its Applications (link) — "Someone who opened an account within the last six months cannot have an average balance (e.g.)for the last year or the last half-year".

Therefore I'm going to slice the history for sequential time points and reweight the slice observations for old and new customers. What's the best way to do this?

Also any literature on creating behavioral models based on logistic regression is appropriate.

$\endgroup$
2
$\begingroup$

There are different ways to tackle the problem.

  • You could just take the average balance for as long as it is available and up to one year. This assumes that recent accounts are as likely to default as older ones.
  • You could take the average balance for as long as it is available and up to one year but also add variables that reflect the age of the account. Those variables could be a continuous variable with the age of every account or it could be a binary that flags only recent accounts (those where you computed the average balance over a shorter amount of time). Your algorithm can then find out on its own whether new accounts are more risky and reflect that in the score.
  • You could delete all recent accounts from your data-set.
    • This will reduce the number of records to learn from.
    • This will make it problematic to predict the score of young accounts if you only had old accounts in the training data.
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.