Looking through a popular public dataset for risk scoring in home credit from a real home credit company, and some of the features included there are absolutely wild:
"Who was accompanying client when he was applying for the loan?" (Unaccompanied, spouse, partner, group of people)
"On which day of the week did the client apply for the loan?"
"Approximately at what hour did the client apply for the loan?"
"Number of elevators in the building where the client *currently lives*"
"Number of enquiries to Credit Bureau about the client one hour/day/month before application"
... I don't doubt these might be *correlated* with defaulting on loans, but someone at these agencies must realize that using these features for individual-level decisions will result in absurd outcomes?
Nothing can beat credit scoring based on writing style though: