Credit scoring in online retail: What your data trail reveals

Even just the handful of data generated by an online purchase is enough to yield a credit score that’s as sound as one calculated by a credit rating agency like Schufa. A research paper by the Frankfurt School of Finance & Management shows how important our digital footprint is.

  • Online retailers can produce a credit score for their customers with just a little bit of information.
  • Information such as type of smartphone used or email provider allow conclusions to be drawn about how likely the customer is to default on payment.
  • Valentin Burg, visiting scholar at Humboldt University, explains how the model works.

Mr. Burg, from the data that every online store collects about its customers your team was able to produce a credit score that is as good as one provided by a traditional credit scoring agency. How did you figure out that something like this is possible?

Valentin Burg: Since completing my doctorate I have been working in an e-commerce company that sometimes gets very imprecise results from the credit rating agencies. This is a problem because we also offer payment on account, which is the equivalent of giving the customer a mini loan. So we asked ourselves: What do we know about those customers who defaulted on payments? What attributes could we use to evaluate creditworthiness? I talked about this to a former colleague who works at the Frankfurt School of Finance & Management and it turned into a research project.

Credit scoring in the online retail segment: Dr. Valentin Burg is a visiting scholar at Humboldt University in Berlin, where he has conducted research into various areas including the influence of management styles on financial decisions.
Dr. Valentin Burg is a visiting scholar at Humboldt University in Berlin, where he has conducted research into various areas including the influence of management styles on financial decisions.

Your model takes account of quite simple factors such as the customer’s email service provider or whether they use an Apple or Android device. What is the reasoning behind this?

We took an intuitive approach and looked at what factors might make sense. For example, there is already some research showing that people with Apple devices tend to have higher incomes. An email address can indicate income too. Can the consumer afford a provider like T-Online or do they use a free service? Other factors are based on behavioral theory: Someone who comes to the online store via an advertising banner is perhaps more likely to be an impulse buyer than someone who comes after visiting a price comparison website.

So do you also check whether your interpretations are correct?

That wasn’t the objective of our research. We are less interested in why these variables play a role than in knowing that they do play a valid role.

But if word gets around that a certain email address or type of phone can affect your credit score then consumers will find ways of getting around that of course.

Our observation is that people are not doing this to any great extent, because after all, there’s a lot of effort and expense involved.

Can you use your model to also establish customer profiles and make long-term predictions?

We looked only at short-term loans that are due after one month. The data doesn’t tell us what someone’s credit standing will be like over several months or years. If I were the operator of an online shop I would do a new assessment each time the customer shops in my store. I would also check regularly to see whether the variables are still valid or if other variables have taken on a bigger role.

This is a very different approach to that of a credit rating agency, as their scores are based on the customer’s history.

Yes, the digital footprint really does represent the here and now. That is probably also the reason why our scores only have a weak correlation with those of the credit scoring agencies – although both are equally precise when it comes to distinguishing between better or worse risks.

Can the digital footprint replace the traditional credit scoring agency at some point?

Probably not. Our paper does show that the two scores complement one another very well. For example, if I have 1,000 customers with an A rating, I can distinguish between those that tend to be better and those that tend to be worse. And if I have customers who get poor rankings from an official credit bureau (e.g. Schufa in Germany) I can find out whether some of them are perhaps actually a good risk and have been rated incorrectly.

Credit scoring in the online retail segment: Completing an online purchase on a laptop.
Every time you shop online you leave a digital footprint behind.

Are fintechs actually queuing up at your door now to incorporate your findings into their own products?

Yes there has definitely been a lot of interest.

From financial services providers and banks as well?

Yes, some of them are interested.

But you cannot say more at the moment.

Yes (laughs).

In a global context you are not the only ones working on digital scoring. The US insurance fintech Lemonade, for example, is collecting data among other things on how much time customers take to read a contract before they click on “accept”.

I think that a lot of companies are using something similar. The best-known example is probably in China, where Alibaba subsidiary Alipay is developing credit scores that are effectively based on behavioral theory models.

Do such approaches comply with the GDPR?

I can’t say much about that; you would have to evaluate it on a case-by-case basis. But any company doing something like that naturally has to be transparent about it.

So that the customer knows why they are not being offered certain payment options.

It is also the case that digital scoring can actually work to the customer’s advantage, if they can use an insecure payment method despite a poor official credit rating. The point is to assess customers more accurately.

When would you advise a company to develop a scoring system based on your model?

It’s a question of scale. Can you afford it? And do you have enough data to be able to make accurate assessments? It would need to be in the region of a few thousand observations per month. In such circumstances I would recommend developing a system in-house. Not just because of the better quality of the lending decision but also because you learn even more about the customer.

Photo credits: E+ / Getty Images (2), PR

share this