32%Real lending data from an open research dataset
What declined applicants would have done
of the error on declined applicants closed, where the answer was known
- What was observed
- 192,699 approved loans with outcomes, and the applications that were declined, which have none.
- What a generic approach says
- Judge declined applicants as if they were like the approved ones.
- What the engine read
- Weighting approved loans to look like the declines closed 32% of the error, on a test where a stricter policy was imposed on loans with known outcomes so the answer was known. It did not rank declines any better, and 84% of real declines were so unlike anyone approved that no estimate was made.
- What it meant for the decision
- Declines can be estimated with a measured correction, never assumed, and a small test that approves some at random measures the rest. The test used only recorded fields, the best case for this method.