I always felt that determining consumer risk for lending is a simple problem.
Consumer lending is a risk-probability based business. To be a profitable lender, one must be able to predict the future. Traditionally, inferring the customer risk was done by looking to the past: The premise that if a customer were good at paying back in the past, they would continue to be good, is generally sound. In developed markets, a credit bureau provides access to this data within seconds. However, the over-reliance on historical data is uncomfortable.
In markets where there is no credit bureau, lenders have utilised a simple model to lend — a personal relationship with the bank, ID check, bank balance, etc. Here, the risk model is not complicated: All borrowers pay the same interest rate regardless of the risk level. If you are an unbanked customer in one of these markets, your choices are slim - ‘Help’ for such customers comes in the form of friends and family or an extortionate, unregulated high-cost lender.
A few years ago, due to the exponential growth of smartphone usage, some start-ups saw an opportunity to create mobile credit solutions for the unbanked and the “thin filed”. This was an exciting challenge; it was an opportunity to develop a better risk calculation model and to automate lending to the core. If a business could lend money based on a statistical model, without the need to manually validate documents, it will have exponential scalability. By design, the operational cost of such a company will remain low due to automation which is a massive advantage).
As a result, many lenders (predominantly start-ups) embarked on creating proprietary solutions to minimise the inherent risk in lending by capturing non-traditional data – such as mobile data, GPS, behavioural data, etc. – as a proxy for the level of risk (or lack thereof).
There is a catch-22 to this approach: to create a risk model one requires data; to know which data to capture, one needs a risk model. Knowing we had to start somewhere, the data scientist created models that were predominantly based on conjecture and logic.
At inception, these scores were not optimised for the real world. Data Scientist needed real-world data from real customers to calibrate the risk score. The standard strategy was to lend (or promise to lend) a small amount of money (sometimes even as low as $5) over a short period to enable the lenders to gather data at a low overall risk to the business.
A simple hypothetical underwriting system can capture data such as:
· GPS data that demonstrates the user has a routine
· AND has more than 150 contacts in his/her contact list
· AND has had a mobile phone for over a year
· AND has between 60-90 apps installed
A lender can conclude that customers who match the above criteria can be lent money. By measuring customer success, behaviour and other traits during and after the loan, data-scientists were able to calibrate risk scoring to an acceptable standard.
It is crucial to remember that the distinction between correlation and causation is frequently misconstrued to convince investors and shareholders of the business opportunity. Using thousands of data points to underwrite a customer where there is a clear correlation, but no clear causality must warrant further research.
Today, these risk models have evolved to consume thousands of data points. They no longer look at data points in isolation: they consider the correlation between the data points, the weight given to them and any exceptions to the established standard rules. Thus, these models have become uber complex. Keeping such complex models relevant requires one to adjust the model frequently to the ever-changing circumstances.
In such a business, a Risk Manager would be unable to explain the exact reasons why the company lend to some and not to others. They are unable to establish the hypothesis for the interest rate or loan value offered to one customer over another.
Since these risk models were not built on axioms but on an arbitrary set of rules that loosely correlate to customer performance, they are susceptible to catastrophic failure. The introduction of stringent data protection rules or some form of disruptive tech innovations that obstructs the access to data would render the risk model obsolete.
In my opinion, risk models could be far more straightforward. There are two questions to answer when you lend money to an individual: one, can they afford to pay it back?; and two, how much do you trust the individual? — essentially, an evaluation of the customer’s affordability and trustworthiness. I am assuming that at this point of the customer journey, the customer has been deemed to be non-fraudulent and internal credit policy compliant.
Trust score, while subjective, can be determined with a level of acceptable accuracy in developed markets. A customer credit history provides the data needed to calculate the trust score. Infusing the past with other data points such as the provision of formal ID scan, an on-the-spot selfie, a scan of the customer’s bank card, and details of the length of time that the customer has had a bank account (among others) can establish an enhanced trust score. With accurate engineering, the score can be calculated with no human interference.
The most exciting development in the UK and other developing markets is the introduction of Open Banking, which automates the computation of the affordability score using real-time, raw data. Open Banking is a game-changer!
In its purest form, a lender can look at the historical bank balances of a customer. If the overall monthly saving (3-12 months) exceeds the monthly loan instalment amount by 75% or more, the customer’s affordability can be deemed substantially low risk.
A simplistic risk calculation could be:
Customer Daily interest Rate = lowIN + (highIN - lowIN) x (1- (tScore + aSocre)/2)
lowIN = lowest interest rate given to a customer
highIN = highest interest rate given to a customer
tScore = customer trust score
aScore = affordability score
Open Banking creates a balanced ecosystem for lenders and borrowers to co-exist. Lenders can accurately predict a customer’s risk score using information that has a strong correlation to customer affordability and will no longer have to rely on the arbitrary models of old.
This transparency is long overdue: Customers will no longer feel anxious about borrowing money; Lenders will be able to create highly automated, scalable products with ease. It’s a win-win.