Machine Learning and the FICO Score
Blog: Enterprise Decision Management Blog
Greetings from sunny (and humid) Miami! I had the pleasure of speaking on a panel at ABS East yesterday, entitled “Traditional vs Non-Traditional Underwriting, Does Machine Learning Teach Us Anything New?”
The panel primarily focused on the opportunities and challenges associated with the use of Machine Learning (ML) in credit underwriting. I called out some highlights from FICO’s recent white paper on this subject. (We will dive into this further in future blogs so keep an eye out.)
On the ‘opportunity’ side, I cited:
- The speed to powerful insights that ML offers, making it ideal for R&D efforts aimed at assessing new analytic challenges, and/or the potential of new data sources to add incremental lift.
- The great strides we’ve made at FICO as far as developing explainable artificial intelligence (AI)/ML and how that enables us to understand better than ever before what the key variables and risk patterns are that are driving the ML-based prediction.
On the ‘challenge’ side, I cautioned that unconstrained ML can lead to counter-intuitive patterns being encoded in the resulting model—a definite ‘no-no’ in the credit risk score space, particularly for a score like the FICO® Score that is used in billions of lending decisions each year, and which receives intense scrutiny from clients, regulators, and consumers alike.
My fellow panelists raised some excellent points as well. My fellow Northern Californian Julian Grey of Black Knight cited the potential for ML to be not just a tool for building highly predictive models, but also for ‘learning’ more effective matching algorithms. She cited success Black Knight has had using such ML-driven matching algorithms to pull together disparate data sources to drive more powerful models. Tamer El-Rayess, CEO and Founder of Continental Finance shared his view that ML offers the greatest promise in prediction problems that do not exhibit standard distributions, due to the fact that ML makes no parametric assumptions about the distribution it is trying to learn.
At FICO, we’re investing considerable R&D resources around explainable AI/ML, and its potential for seamless application to predictive model development. We will continue to test and learn on cutting-edge ML approaches. Specifically, within credit risk scoring, we’re seeking to strike a balance between the power and speed of insights that can be derived from ML and our 25+ years of domain expertise in the field. Ultimately, our aim is to ensure that the resulting FICO® Score model is a predictive as possible, while still adhering to our high standards of model transparency and palatability.