Re:Max Levchin’s Affirm Raises $275M to Make Loans
by derefr on Wednesday May 06, 2015 @11:20PM
(#9503119)
Presumably, Affirm is an application of the same "all-encompassing many-sources-as-possible Bayesian-analysis engine" behind both Paypal's anti-fraud and Palantir's anti-terrorism technology. As such:
> Also, is there data that shows that someone with a long commit streak is more creditworthy than someone without a Github profile?
That data doesn't exist; the precise job of this sort of correlation engine is to create it. In credit analysis, you start with a decent credit-worthiness model of all your customers—derived not from predictions, but from how people with a given feature-set did historically. Then you go looking for new feature datasets to incorporate into your model which decrease its RMSE at predicting split-subsets of your outcome data.
Github commit history could easily be one of these. I would guess it would be a good prediction of both average employment tenure, and of the Big 5 Conscientiousness personality trait.
But I don't have to be right—I just throw the feature-set into the engine, and it either finds signal in it and turns up its weighting in the model, or finds that it's noise and turns it down to zero.
> Also, is there data that shows that someone with a long commit streak is more creditworthy than someone without a Github profile?
That data doesn't exist; the precise job of this sort of correlation engine is to create it. In credit analysis, you start with a decent credit-worthiness model of all your customers—derived not from predictions, but from how people with a given feature-set did historically. Then you go looking for new feature datasets to incorporate into your model which decrease its RMSE at predicting split-subsets of your outcome data.
Github commit history could easily be one of these. I would guess it would be a good prediction of both average employment tenure, and of the Big 5 Conscientiousness personality trait.
But I don't have to be right—I just throw the feature-set into the engine, and it either finds signal in it and turns up its weighting in the model, or finds that it's noise and turns it down to zero.