Where’s the Watson of credit risk?

If the number of memes on Facebook are anything to go by, most of us are hanging out for 2016 to be over. It’s been a chaotic year across the globe – we’ve seen political upheaval in two of the world’s biggest economies, lamented the deaths of creative geniuses David Bowie and Leonard Cohen and witnessed mass displacement of civilians as a result of escalating conflict in the middle east.

As if that wasn’t enough, it hasn’t been a great year for small business online lenders either, or their backers. LendingClub’s stock is down 51 percent on the year, while On Deck Capital has fallen nearly 60 percent since January. Prominent merchant cash advance provider CAN Capital was also reported to have run into trouble with its repayment collections process, as growth in originations outstripped the ability for internal processes to keep up.

In other news, reports are surfacing that mid-prime lender DealStruck has potentially closed its doors on new business.  Crowdfund Insider reported the closure followed a failed acquisition by an unnamed ‘Utah based bank’. Ex Chief Strategy Officer Candace Klein provides an interesting autopsy here on some of the possible industry drivers for the closure, touching on the increasing competition amongst new lenders and banks for prime and super-prime borrowers, leaving the ‘too hard basket’ of the mid-prime borrowers out in the cold.

While I’m not privy to the inner workings of DealStruck, or necessarily why they decided to shut up shop, I can’t help but wonder if, like the famous Greek Sirens, the lure of lending to the underserved mid-prime business market is a shortcut to shipwreck for many? It seems even the biggest non-bank lenders with arguably the deepest pockets are still struggling to develop scalable, repeatable and dependable credit risk models that can help them scale safely and price effectively. Maybe it really isn’t as easy as everyone had hoped to turn shades of grey into black and white decisions, without human input.

Can the art of credit ever be turned into a science? Scores of fintech lenders depend on the answer to this question being an emphatic yes. But this relies on credit risk models being able to learn more adaptively, as humans would. Finding the sweet spot between the push and pull of quantitative computerised models and qualitative human based decision making is the nut that many online lenders are yet to adequately crack. We need a machine learning approach here, that moves beyond a reliance on inputs only, but learns from the mistakes and the successes to build lookalike credit decisions.

Supposedly many online lenders use machine learning today.  I’m sure there are also a large number that still rely on classical, human driven decision making, hidden beneath a slick mobile interface. While poorly executed machine learning could be worse than human driven decision making, well executed machine learning is certainly a winner on the scalability front – and that is the prize in mid-prime lending.

Daily Fintech Advisers provides strategic consulting to organizations with business and investment interests in Fintech. Jessica Ellerm is a thought leader specializing in Small Business.


  1. I think credit risk is a science already! Many banks have been using credit models successfully for over 20 years. However, it seems to me that scale is a key ingredient. Without it, you are at risk of losing all of your profit from a few bad loans. In my view,we will struggle to reliably predict the future probability of default at the level of an individual firm. But we can predict the future with great accuracy for a large enough population.

  2. There are three issues of credit risk adjustment here. Collection of evidence, learning from evidence and adaptation to evidence. Collection of evidence machine or human depends on the availability of the appropriate data and this is about business transactions and not strictly financial that most lenders collect. Learning of evidence and identifying pattern of performance human are still better than machines. On the other hand, machines adapt more easily to evidence learned as the process is more automatic and they do not suffer from hysteresis of sentiment, memory and inertia!

  3. The best model in the world will never turn a poor lending proposition into a good one. Yes, there may be marginal gains from improved models but I cannot see anything on the horizon which will substantially change the good:bad odds in this segment, which is the heart of the issue. We are currently at a benign point of the economic cycle – the big question is what the scene will look like when the tide goes out.

    • Its all about the data I think. If you have data as fine-grained as say Asian restaurants by price point by by socio-economic neighborhood (to take one example) you can see how they fared in the last macro cycle. I do think credit decisions will be augmented intelligence ie humans with better data and models not artificial intelligence (ie no humans).

  4. For all but the smallest business loans, I think the key lies in operating models that deploy expensive human expertise cost-effectively, so that alternative lenders can compete in segments that traditional banks – with expensive physical presence and inflexible legacy systems – find uneconomic.

    Examples of this could be using modern data collection techniques to get business performance data (e.g. connecting to accounting software APIs), or taking the paperwork out of AML/KYC by deploying remote mobile online authentication tools. Expensive staff can then be centralised, and focused purely on relationship-management and decision-making.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.