iPhone 6s
ƒ/2.2
4.15 mm
1/30
80
Apaar Sadhwani of Google Brain opened the ICME AI Forum with a summary of their work to use simple neural networks (only 5 layers deep, 100K parameters) to outperform all of the logistic regression techniques standardly used to predict default and pre-payment risk in mortgage portfolios (Logit in red here).
The model benefited largely from ingesting local risk factors at the zip code level.
He trained the network with 70% of all mortgages over a 10 year period.
From the AI in Fintech Forum, hosted by the Stanford Institute for Computational & Mathematical Engineering (“We do big math”).
Relative importance of input variables:
A lot of cool ideas but for regulation:
Leave a Reply