Revolutionizing China’s Consumer Lending Industry with Machine Learning
Machine learning and big data are used by online micro-credit providers to grow China's burgeoning consumer credit industry at record pace. Will these new tools introduced in recent years create or destroy absolute value in the long-term? How will the leaders of the industry leap forward in an increasingly short innovation cycle?
In the Learning and Innovation module, we briefly looked at how companies applied machine learning algorithms to solve challenges that once seemed impossible to accomplish, e.g. IBM Watson takes on Jeopardy! [1], and Aspiring Minds of India tries to solve talent market inefficiencies [2]. In this article, I examine how machine learning disrupted the micro-credit lending market in China, while drastically lifting the productivity and diversity of the financial ecosystem.
I view the evolution of consumer credit industry in four phases:
In Phase One, there is no centralized borrower of scale, and people borrow from people they knew, or worse, from loan sharks within distance they can reach.
In Phase Two, companies use a dozen or so inputs to check the boxes on a customer’s application. If all boxes are checked, a loan is approved. A number of banks still make lending decisions this way today.
In Phase Three, a credit scoring system is introduced. I will use FICO score as example. In the U.S., consumers closely monitor their credit/ FICO scores, as it can have a direct impact on the annual percentage rate (APR) they get on their auto, home mortgage loans, as well as credit card or loan application successes. FICO, a credit score provider, doesn’t control any consumer data itself, but offers services to help banks, insurance companies, etc. to build their in-house personal credit risk assessment systems. A look at FICO’s products shows Analytic Consulting is just one out of near a hundred products it offers to businesses [3].
We are in Phase Four of consumer lending, where companies use machine learning and AI to come up algorithms that automate the decision-making process. For example, Qudian Inc. of China – an online small consumer loan provider, was three years old when it went public on NASDAQ in October 2017. The company capitalized on the wealth of consumer data that became available in the mobile internet age, and used machine learning capabilities to 100% automate its lending decision.
Whereas a single credit score based on credit report pulled from People’s Bank of China (China’s Central Bank) is used in Phase Three, data inputs from a myriad of sources are used to compute multiple scores that inform the lending decision. It’s a black box where the machine assigns weights to data inputs and spits out a decision. Feedback is a vital step in the process –a consumer’s behavior after the loan is taken out is fed back into the machine learning algorithm to refine and improve the algorithm, so the predictive power of the algorithm on how creditworthy and profitable each customer is gets stronger over time.
What Qudian and its online micro-leading peers addressed was hugely unserved or underserved demand for personal credit in China. Compared to the U.S. where the average number of credit cards per person was 3.7 in 2014, China’s figure was 0.34 [4]. Out of the 1.4 billion Chinese population, only 200 million had access to credit cards. Many people either lack credit history or geographic proximity to a bank branch, and therefore could only resort to loan sharks to meet financing needs. This is why the impact of Qudian’s innovative product – a credit line whose entire application and approval process happens online, all determined within a day after the customer submits basic identification information – is revolutionary for China’s financial system [5]. As of August 2018, Qudian has made over 140 million transactions [6].
Source: QuDian Inc. IPO Prospectus [7]
Another interesting feature of online consumer lending enabled by machine learning was dynamic pricing – the algorithm will not only decide if a loan can be issued, but also at which APR the offer should be made. This feature wasn’t possible before, because the computation required for setting APR would further add to the time it takes to make a loan decision, reducing efficiency in an already time-consuming process. The algorithm is built on the computing power of machine learning to accurately predict the estimated cost associated with making a loan (including bad debt cost, which is the write-off the lender has to make when the loan deteriorates in quality and hits a pre-determined time overdue limit). The ability to process and analyze big data to offer a personalized APR means the number of borrowers whose credit needs are covered is maximized. Whether a company has a good machine learning algorithm that constantly reiterates and improves will determine that company’s profitability.
Where the industry goes next will be interesting. My big question is about the superiority between supervised and unsupervised learning in a lending decision context, and how to apply the findings to further disrupt the financial services industry in a value-additive way?
(787 words)
Sources:
[1] Harvard Business School Publishing, “Building Watson: Not So Elementary, My Dear! (Abridged)”, HBS No. 616-025, October 21, 2016
[2] Harvard Business School Publishing, “Aspiring Minds”, HBS No. 616-013, May 6, 2016
[3] FICO Products. https://www.fico.com/en/products/fico-analytic-consulting, accessed November 2018
[4] Gallup, “Americans Rely Less on Credit Cards Than in Previous Years”, April 25, 2014, https://news.gallup.com/poll/168668/americans-rely-less-credit-cards-previous-years.aspx, accessed November 2018
[5] Seeking Alpha, “Qudian: The Risky Chinese Fintech Opportunity”, June 1, 2018, https://seekingalpha.com/article/4178789-qudian-risky-chinese-fintech-opportunity, accessed November 2018
[6] Qudian Investor Relations, Qudian Inc. Provides Business Update, http://ir.qudian.com/2018-08-03-Qudian-Inc-Provides-Business-Update, accessed November 2018
[7] SEC, Form F-1 Registration Statement, https://www.sec.gov/Archives/edgar/data/1692705/000119312517287443/d282719df1.htm, accessed November 2018
Dynamic pricing’s ability to get less creditworthy borrowers loans that they would otherwise not have had is very powerful (though the higher rates may be a concern for less sophisticated borrowers). I wonder what kinds of biases these algorithms may have. In particular, I wonder if there may be a similar issue as seen in Amazon’s recruiting algorithm, where the heavier proportion of male to female skewed success rates towards males, may also be present for demographics that have been less represented in certain success areas. This would be a key area of focus in developing the algorithms to be agnostic to demographic biases.
I wonder what type of forward-looking capabilities the algorithms have. QuDian may maximize current profits by reducing loan/bad debt costs, however the risk and expected returns change as the competitive landscape evolves. As more competitors emerge and create similar offerings, QuDian and other lenders need to generate greater differentiation, rather than iterate and evolve in the same manner. Perhaps Phase Five of consumer lending takes supervised learning and lender feedback into greater consideration.