Machine Learning for Finance
A Practitioner's Guide to Building, Validating and Defending a Credit Model That Decides
-
- $12.99
-
- $12.99
Publisher Description
Five people quote the accuracy of the same credit model and give 0.9125, 0.7353, 0.7375, 0.7185 and 0.6984. Nobody is wrong: each is the correct answer to the question its author was asking, and the distance from the highest to the lowest is 0.2141 of area under the curve - most of the distance between a good model and no model at all. Brechin Credit is an unremarkable lender to British small companies and its model is an unremarkable model: seven characteristics, weights of evidence, a logistic regression, points on the industry scale. The differences come from one field that is populated three months after the money is lent and takes 0.1772 with it when it goes, from a year the model had not seen, and from the fact that the only cases with an outcome are the ones a previous policy accepted. And the number that decided how much money the lender made is none of the five. It is 8.2367 per cent, the bad rate at which a marginal loan is worth writing, being 6,982 of margin on a performing loan divided by the 84,767 swing between a performing loan and a default. That ratio puts the cut-off at 550 points, approving 83.50 per cent of applications and earning 17,348,284; at 600 points, the score a nervous committee reaches for, the same book approves 16.80 per cent and earns 5,596,817. Carefulness costs 11,751,467. Then 2020 arrives: the ranking barely moves, the level under-predicts by up to 9.612 percentage points, and the profit-maximising cut-off moves thirty points to 580, so a policy written once and left at 550 turns 2,563,882 into a loss of 149,416. One dataset of 8,400 applications carried the whole way through, with the arithmetic printed rather than asserted: the outcome window, binning and weights of evidence, the points scale and what each characteristic is really worth, discrimination against calibration, population stability, the applicants the model never saw, the cut-off priced in both years, whether 180 boosted trees earn their keep on the same seven inputs, explaining a decision to an applicant and to a regulator, the monthly monitoring pack, and the seven questions to ask of somebody else's model paper. Four companion workbooks reproduce every printed figure by a live formula and check themselves against the page. For credit risk modellers and validators, model risk and internal audit, chief risk officers and credit committees, and data scientists arriving in a regulated lender from somewhere else.
About this edition: the text of this book was generated with AI (Claude) and published by the author.