Quote:
Originally Posted by Wolf Larsen
If I were going to do this, I'd collect a bunch of point in time social data and loan history data and then use it to train a neural net to figure out which things were predictive of repayment problems and what the time lag was. I'd then validate it with some new data to make sure it still had predictive value before I let it be a weight in the underwriting equation.
|
This is my biggest issue with Big Data. There's this assumption of predictive value that comes along with these ideas of using social and behavioral data points to find most desirable/least desirable customers.
Vetting shit like this is part of my job. Ninety-nine times out of a hundred, these Big Data companies haven't validated their supposed predictive abilities with any hardcore evidence. Makes for some interesting meetings once you mention this, especially after having asked some fun questions about data-gathering techniques and their compatibility with established privacy policies (and the prevailing attitude toward snooping...)