0
$\begingroup$

I would like to build a scorecard from scratch.

Currently I have been working on data cleaning, and I created a dataset for training and another for test. What I have is a dataset of around 700 rows (maybe not so many as I could expect; it is for a computer science project), 20 columns. The target is if the borrower has defaulted or not (indicator). Most of the factors are categorical or Boolean. I am assigning a score based on these parameters, specifically:

- if the account has defaulted in the last 90 dats, then assign 60;
- if the account has defaulted in the last 30 days, then assign 50;
- if the account has not defaulted in the last 90 days, then assign 80;
...

I would like to know if this is a good approach and/or if there is an algorithm to assign scores to the accounts or if I should do something different. I do not know very well the logic behind a scorecard, so any information would be extremely useful.

Thanks

$\endgroup$
0
$\begingroup$

What you describe it is NOT a typical approach for scorecard design.

Typically binary target - "0/1" is used for target variable and logistic regression is typical method to create a scorecard.

Typically to define "default trigger" one fixes two time periods:

1) fix say 1 year from credit start, that means defaults happening after 1 year are not considered as defaults

2) fix say 90 day period - means "default trigger" is 1 if client delays payment for more than 90 days from the payment date, that means you do not consider as "default" if delay is less than 90 days.

There can be some other conditions - like threshold on amount of delinquency - if below - do not consider as "default". And may be some other.

You might consider the book:

Credit Risk Scorecards: Developing and Implementing Intelligent Credit Scoring

by Naeem Siddiqi

$\endgroup$
5
  • $\begingroup$ Thank you so much for your answer Alexander. I would have a further question: in terms of logistic regression, what information would I need? could you please give me an example more practical? I will read the book you suggest me. Thanks a millioon $\endgroup$
    – Math
    May 10 '20 at 20:46
  • $\begingroup$ @Math you would need your dataset and target variable (binary target default/non-default). After that you can try to build a logistic regression. It takes form P (defaullt ) = 1 / (1 + exp (linear_function(data)) . That "linear_function(data)" is "score". "Logisttic regression" in sklearn-Python/R/... return you that linear function, when you give data, target as an input. Typically one makes a preprocessing of features - splitting them to "bins" and substituting by WOE transform, but that might take us too far away from the question... $\endgroup$ May 10 '20 at 20:55
  • $\begingroup$ Thank you so much Alexander. I will have a look at the book and at the sklearn toolkit. I would need to build a scorecard from scratch and since I have never used it before, I wanted to understand what there is behind, step by step, through an example. Do you have any idea where I could find a really small example (just with a dataset of a few rows - but enough to build it ) to understand its logic? Even outside the credit scoring card. I think the most important thing would be to apply a logistic regression model and then assign scores.. $\endgroup$
    – Math
    May 10 '20 at 21:15
  • $\begingroup$ @Math first - it is typical binary classification problem - so any binary classification would suffice to master python/sklearn/logistic regression, there are plenty of them e.g. Kaggle provides lot of. Second - there are specific details related to business - you might look at the Siddiqi book which is somewhat standard reference $\endgroup$ May 10 '20 at 21:21
  • $\begingroup$ Thank you so much. I will have a look at Kaggle repositories and at the Siddiqi book. $\endgroup$
    – Math
    May 10 '20 at 22:59

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.