How FinTechs are rewriting underwriting?

Of the over 2000 FinTechs in India, only 4 find a place in the list of 20 plus Indian unicorns and none of them are in the lending business. There is something about lending that goes beyond technology; one, it requires large capital and continuous fund infusion and two, there is substantial risk which tech cannot fully address. What brought FinTechs in droves initially were two factors- the enablement of e-KYC using Aadhar/India stack, which revolutionized customer onboarding, and the huge unmet credit market that traditional banks had ignored (because they were new to credit and lacked conventional credit rating).

FinTech lending is relatively new and start-ups entered only around 2013, therefore it is still early to judge. The initial focus was on outputs – faster TAT (loan in 30 minutes), zero or limited human interaction and improved customer experience. With these low-hanging fruit having been plucked, the differentiators will be in risk and credit underwriting which is where AI and ML are being increasingly utilised. The risk assessment and underwriting journey itself has been fascinating. Even globally, the underlying theme has been that that the traditional three-digit credit score is not onlyan incomplete view of peoples’ ability to pay back but also shuts out millions of deserving borrowers. Enter fair algorithms which can provide a more nuanced and accurate assessment of an individual’s risk profile.

OnDeck, a leading US based digital lender has its proprietary tool OnDeck Score look at variables such as number of customers, cash flows, sales, registered complaints, vendor payment history etc rather than only tax returns or financial statements.

Since they chose to focus only on small businesses with steady cash flows, such as dentists, restaurants or repair shops that had a trail, they were constantly scouring “the cloud” to develop a clear understanding of a potential client’s creditworthiness. For a restaurant, for instance, they even look at sites such as Yelp-a large number of reviews tells them more about revenues or quality than financial statements could.

The above is still a refinement of traditional credit assessment, but it is in the area of alternate data and application of AI and MLthat credit evaluation gets interesting. Much of this seems to beunsupervised AI, where humans don’t create rules at the beginning but let AI identify patterns across millions of variables fed in.The number of data points collected are usually massive (tens of thousands) and even bizarre. For instance,Lenddo provides a social credit score that predicts an individual’s ‘willingness to pay’, based exclusively on his social data and online behavior.  Their algorithm considers features that we may consider weird- such as avoiding usage of one-word subject lines (you care about details) or the number of financial apps on your smartphone (you take your finances seriously) and even the ratio of smartphone photos taken with a front-facing camera (to segment customers into young and old!). Tala is another data science based lender that relies on unconventional sources, mobile usage data in this case, to evaluate customers. They too pick up odd data on grammar, punctuation or the time of day when most calls are made, or the number of contacts and duration of calls (for instance calls to more than 58 people and an average call duration of 4 minutes or more is said to indicate ‘strong social network’ which is reckoned as good for repayment!) and crunch them to come out a financial score in under 10 seconds and money is disbursed the same day.

In India, digital data availability is a major limitation which is probably why we don’t find many robust use- cases in credit underwriting. One of the bigger players Capital Float, is stated to extensively deploy AI and ML models; reportedly, their kirana app uses a roundabout way (using data from sales of prepaid mobile recharges) to construct a credit profile for the kirana shop as the shop itself has no digital footprint. Another, LendingKart, is also said to use AI to crunch more than 8500 data points on the customer (not from past financial statements or tax returns but non-conventional sources). But there isn’t enough evidence to show these models are working given the short period they have been in operation although bad loans ratios have been lower than traditional banks. But more than data availability, the bigger problem with AI and ML will be ‘explainability’- users need to be able to explain their “black-box “decisions. This is still some distance away for India, but in countries like the US, legislation such as the Fair Credit Reporting Act of 1970 requires accurate and actionable reasons for credit denial decisions so that consumers can repair their credit and re-apply successfully.