Score an applicant and see exactly why.
A retail credit scorecard trained on 60,000 synthetic loan applicants. Change any detail and the score, the decision and the reasons behind it update.
The scorecard
Each row is one factor, split into ranges worth a fixed number of points. The highlighted range is where this applicant falls, and the highlighted points add up to the score. Factors are ordered by how much they separate good loans from bad ones.
How it's built
The standard method for application scorecards at retail lenders, chosen because every decision can be explained to the applicant, an auditor or a regulator.
-
Bin each factor
Each input is split into ranges, then neighbouring ranges are merged until risk moves in one direction only. Each range gets a Weight of Evidence, the log ratio of good to bad loans in it. Factors that barely separate the two are dropped.
-
Fit the model
A logistic regression is fitted on the binned factors, using only the training data so the test results are honest. Six of fourteen candidate factors were dropped for weak signal.
-
Scale to points
The model is converted to points: 600 at odds of 20 to 1, and every 50 points doubles the odds of repayment. Scoring by points gives exactly the same ranking as the model.