I'm looking for recomendations of books on credit scoring. I'm interested in all aspects of this problem, but mostly in: 1) Good features. How to build them? Which have been proved to be good? 2) Neural networks. Their application to credit scoring problem. 3) I've chosen neural networks, but I'm interested in other methods as well.


If you are new to the scoring world, your first book should be by naeem siddiqi on credit scoring using SAS. If you have not taken the class go for it. The class main focus is the overall understanding of scoring and selling SAS enterprise miner for millions of dollars.

If you need theory you need a categorical data analysis and Data mining class from a near by university. Even after taking these classes you will still need help.

currently the most popular techniques used are

  1. logistic regression
  2. neural networks
  3. support vector machines and
  4. random forests

clustering, discriminant analysis, factor analysis, principal components are a must as well.

Credit scoring by elizabeth mays will also give you a good overview.

I also took a credit risk modeling class by SAS institute, which helped me a little. It is a constant learning process and its never done.

Bayesian folks like their methods as well.


i also forgot to mention. Logistic regression in the most popular technique out there and will always be the one that banks will continue to use. Other methods are very difficult to sell to the upper management people, unless your bank is willing to care less about understanding these methods and their focus remains risk taking and money making.

  • $\begingroup$ Thanks! I'm going to clarify: I'm participating in an online competition, where the goal is to predict the probability a loaner's default. So 1) I'm free to choose any method I like. The competition ends in 2 weeks, 2) I don't have a lot of time to do comprehensive learning. and 3) The provided data is raw responses from credit bureaus on previous credits, so I am really interested in extracting non-obvious features from this data. $\endgroup$ – Nya Apr 5 '13 at 15:10
  • $\begingroup$ Also, thank you for your response, I'm definitely will take a look at your references. $\endgroup$ – Nya Apr 5 '13 at 15:10
  • 1
    $\begingroup$ What is that competition? Can I know? $\endgroup$ – xiaodai Aug 15 '13 at 13:55

I work in the credit scoring field. Even though I like exploring different approaches I find that logistic regression is often good enough if not the best approach. I have not surveyed the most recent papers on the topic but from memory in most papers you will see that other approaches such as neural nets model typically do not offer significant lift in terms of predictive power (as measured by GINI and AR). Also these models tend to be much harder to make sense of to a layman (often the most seniors executives do not have backgrounds in statistics), and the scorecard approach using logistic regression seems to offer the easiest to explain models. True, most scorecards don't take into account interactions, but again there is no study in the literature that can clearly demonstrate that incorporating interactions consistently and significant increase predictive power.

Having said that, there has been recently some interests in building scorecards using survival analysis techniques as it holds a few advantages over logistic regression. Namely, we can more easily incorporate macro economic factors into the model, we can use more recent data in the model build instead of having to rely on data at least 12 months ago (as the binary indicator in logistics is usually defined as defaulted within the next 12 months). In that regard my thesis could offer another perspective in that it explores building credit scorecards using survival analysis. I showed how survival analysis scorecards "look and feel" the same as logistic regression scorecards, hence they can be introduced without causing too much trouble.

In my thesis I also described the ABBA algorithm which is a novel approach to binning variables.


Update: I make no claim to whether my thesis is any good. It's just another perspective from a practitioner in the field.

  • 1
    $\begingroup$ This does not seem focused on the question. Your concrete suggestion is "read my thesis". I haven't read it and am not qualified to evaluate it, but it doesn't qualify as a book, or even a published paper. $\endgroup$ – Nick Cox Aug 15 '13 at 10:33
  • 4
    $\begingroup$ @Nick The question does ask for "other methods," which this answer addresses. Many people find a link to a downloadable text to be useful, perhaps even more so than just a reference. To an anonymous flagger: providing a link to one's own work is not spam. We welcome researchers and other innovators here and would not want to limit their abilities to help us by requiring that they never cite their own contributions! $\endgroup$ – whuber Aug 15 '13 at 13:58
  • 3
    $\begingroup$ I take @whuber's point. I also strongly agree that citing one's own work is in order. xiaodai: I'd remove the Update. The whole point of your post is that your thesis may be worth reading. If you didn't think so, you wouldn't post. Adding a note of diffidence or modesty is not needed. $\endgroup$ – Nick Cox Aug 15 '13 at 14:21
  • I have referred to Guide to Credit Scoring in R by D. Sharma in the past and it is a good introductory reference on approaches including logistic regression and tree based methods
  • The above guide uses the German Credit Data which has a rich set of features. If you search for the dataset, you will find other alternative approaches, analysis, and comparisons that may help inform feature selection and model choice for your dataset
  • Neural networks is a fair choice for a binary classification problem as this one. In the real world, a credit scoring model is also expected to provide reasons for why a loan application (say) was rejected. Therefore it helps to have a model where you can identify what features in one's credit history result in a low credit score and cause an application to be denied. Features in regression and tree based approaches are easier to interpret compared to neural networks. If you are evaluating purely on fit, NN is worth a try

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.