Bayesian Networks

Credit Risk Assessment

Bayesian network model for credit risk prediction using Hill-Climbing algorithm and BICScore method.

Course: Introduction to Data Science

Objectives

1Identify risk clients when granting loans using Bayesian Networks.
2Find optimal structure of the Bayesian network using Hill-Climbing algorithm and BICScore method.
3Manually improve network structure so that it is consistent with the business context.

Historical credit evaluation is the strongest risk predictor: ECH=0 (Very Poor) yields 64.5% high risk probability vs ECH=4 (Excellent) at only 18.8%.
The connections found by Hill-climbing with BICScore make sense in the data context, though some causality directions are reversed.
Loans with duration 24-72 months have 52.2% high risk probability, significantly higher than shorter terms.
Automated Bayesian network construction is a good starting point but must be analyzed under business experience.