Here’s why AI won’t approve your loan
Posted: 4 July 2017 | By Charlie Moloney
You’ve just declined your customer a loan, they’ve been deemed uncreditworthy. Why? Because an artificial intelligence (AI) system advised you they would be likely to default. You’ve deferred to the ‘superior’ decision making power of the machine algorithms and big data your AI has at its disposal, but how are you going to explain that decision?
You might be in a board meeting with senior stakeholders, and they might not be impressed to hear that you’ve blindly followed the orders of a bot in the making of company decisions. You’re going to need a bit of insight into how your bot thinks.
Tom Singh OBE, a leading British entrepreneur and founder of fashion retailer New Look, has made a major investment in a machine learning (ML) software company, which helps the financial services and insurance industry make fast, informed, data-driven decisions.
“Logical Glue really bridges the gap between data science and the boardroom”, said Singh, who has a net worth of £385 million, and also backed Veeqo, a Welsh start-up which has created an all in one software for retailers, in 2014.
A notable feature of Logical Glue is their patented ‘white box’ technology, as opposed to the more commonly used ‘black box’, which fully explains the reasons why the ML system makes certain predictions.
A look inside the ‘white box’
“It will give you a score back for its predictions in the same way that any other machine learning or statistical method would do”, Robert De Caux, Logical Glue’s Chief Product Officer told Access AI.
“But it will present the decision in a form that shows you how it was weighed. If we take the example, in the lending business, of credit worthy or uncredit worthy: it will show you that it’s weighed, for example, 32% on the creditworthy side, 68% on the uncreditworthy side.
You’re going to need a bit of insight into how your bot thinks.
“Then it would have two columns effectively showing a list of reasons for why that is the case, and those reasons would be ordered according to how strongly they’re influencing the decision, and they would be in a format interpretable for humans”.
Logical Glue say that lenders using their software are increasing acceptance rates by 40%, recovery collection rates for debt collectors by 18%, and profits by typically 5-20%, whilst decreasing default rates by 15%. 20x less intervention by underwriters has been required.
“Where typical ML output would struggle”, De Caux explained, “is that you could put a factor in and see how it would influence the probability of whether, for example, a debtor will repay or not, but it’s not always clear at an individual level how much that factor was contributing relative to all the other things about that person. That kind of relative importance comes through in the white box model”.