There seems to be a number of startups (Zest Finance, Credolab etc.) that provide credit scoring schemes that rely exclusively on alternative data without considering users historical financial data at all. For instance, Singapore-based Credolab uses thousands of characteristics from a user smartphone (time spent surfing the Web at night, the number of long incoming calls during work hours etc.) in order to estimate user's default probability, while not including any finance-related variables like the frequency of late payments, previous bankruptcies etc. From what I've gathered they probably use some sort of unsupervised learning (some variant of clustering) to group similar users into clusters. However, what's still not clear to me is the following: how do they determine default probabilities for each of the clusters?

Note: I'm aware that their algorithms are proprietary - I'm just interested in finding out how to determine default probabilities for each of the clusters when there's no data corresponding to users' derogatory financial events.

  • $\begingroup$ They probably use prior knowledge or historical data to assess a probability for each group, something like semi-supervised learning. $\endgroup$ – user2974951 Jan 23 at 15:07
  • $\begingroup$ Might be, but I find it difficult to believe they have historical data corresponding to say "users who spend >2h per night surfing the web". $\endgroup$ – BGa Jan 23 at 15:38
  • $\begingroup$ just because they give predictions based on alternative data doesn't mean that's what was used to train the model. ie you could use a model that trained to predict frequency of late payments, previous bankruptcies etc off time spent surfing at night. $\endgroup$ – seanv507 Jan 23 at 17:40

Your Answer

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

Browse other questions tagged or ask your own question.