But lending has historically been a very difficult business for startups, and for good reason:
- It is highly regulated, which creates significant cost overhead, and the consequence of a “foot fault” is catastrophic. “Move Fast and Break Things” worked well for Facebook, but people are not very understanding if you can’t tell them where their money went.
- Many areas of lending underwrite based on industry standard scores such as FICO. This has made underwriting a commodity, which has led to competition being based on low costs, and on brand. In both cases, big banks have an advantage over startups.
- For a lending business to grow, it needs access to capital that it can lend. Typically a lot of capital. This can be, at worst, impossible for startups to raise, and at best, highly dilutive.
So what is different now? Why are companies having success in in vitro fertilization lending, and across a wide range of other categories of lending, ranging from payday loans to subprime auto lending to small business lending. This boom in so called “alternative lending” is being driven by a confluence in three factors:
- In the wake of the credit crisis, the big banks are skittish. They are avoiding anything that feels risky after having taken a pounding in subprime mortgages. Many people on the lower end of the credit spectrum simply can’t find credit from banks and traditional sources.
- Low absolute interest rates are mitigating the “cost of capital” advantage that big banks have over startups.
- A proliferation of data sources, and the emergence of big data analytic tools, are enabling new underwriting models.
Hinting at the originality of the site’s data capture, Accel’s Sonali De Rycker, now on Wonga’s board, says: “They use a lot of social media and other tools on the internet you don’t even think about. That’s where the magic is.”
The crux of the algorithm is less about the individual pieces of data — your postcode, the colour of your car, how large your mortgage is — but how these pieces of information relate to one another. Crucially, the data points are stacked against the other pieces of information gleaned from past Wonga clients. By the time Accel came aboard in 2009, Wonga had issued 100,000 loans. That’s 100,000 data sets contributing to an ever growing net of information, and each comprising 6,000-8,000 pieces of information about a borrower. “You build the story by joining up lots of data,” Damelin says. “We pay for that data, but we need it. It’s about computing thousands of combinations to look for things that look wrong – or right.”
This big-data, machine-learning approach to underwriting works for a startup like Wonga because payday lending has a few important characteristics:
- Each individual loan is small. Otherwise, too much capital would be required to run experiments. It’s much easier to run an experiment on a £100 payday loan than a $250,000 mortgage.
- Bad loans fail fast. The average length of a Wonga loan is a few weeks. That allows you to quickly iterate your underwriting model. If you were writing 30 year mortgages that may not cure for years, it would take too long to loop feedback into your model to improve it.
- Default rates are high. The state of the art in underwriting payday loans is not very sophisticated. If you walk up to a payday lending storefront with a pay stub and a drivers license, you can walk out with a loan. You don’t have to be perfect to do a better job of underwriting in this market
- Interest rates are high. This allows startups, with a high cost of capital, to make loans profitably. It also helps that most banks and other lenders with low costs of capital have traditionally avoided this market due to reputational risk.
Because of these factors, Wonga has built a better underwriting model that helps it tell the difference between a good borrower and a bad one. This means that it can charge lower interest rates than other payday lenders. In a commodity market like lending, the lowest rates allowed Wonga to quickly gain a significant share of the market. Wonga is reputed to be on track to do over $500M in revenue this year at a healthy profit margin, in just its fifth year of business.
Building a big data machine learning underwriting model is hard, and it is also expensive. You train your model by making bad loans, and looking for patterns to avoid making similar loans in the future. But learning involves making bad loans; it is the unavoidable tuition cost for teaching your algorithm. Wonga saw 50% default rates when it started.
But now we have existence proof that this problem is solvable. And Wonga isn’t the only company that has solved it either.
In the UK, Global Analytics, operating under the Lendingstream brand, have built a terrific business doing small dollar installment loans, competing in an adjacent market to Wonga. This is a team that came out of HNC, the company that built the banking industries first fraud detection software back in the 90s. They know their analytics.
At Lightspeed, we invested in Zestcash to make small dollar installment loans to the credit challenged in the US. When we met the founders, Douglas Merrill, ex CIO of Google, and Shawn Budde, ex head of subprime credit cards for Capital One, we recognized a team who could build a better big-data enabled machine-learning underwriting model, and they have. Zestcash raised an additional $73M to scale the business in the US last month.
Another notable company taking a similar underwriting approach to a different market is Klarna, which provides a payment system for European ecommerce companies, essentially underwriting credit risk between time of purchase and time of payment. Klarna just raised $155M to expand.
Pawngo is taking the pawnbroking industry online, lending to consumers against high value items like jewelry and premium watches.
With all these companies seeing success, why do I think that there is still an opportunity? Let’s describe all of these companies next to each other, with a couple more notable companies added to the mix:
- Wonga: Subprime, unsecured, bullet loans to consumers in the UK (and increasingly other countries).
- LendingStream: Subprime, unsecured, installment loans to consumers in the UK.
- Zestcash: Subprime, unsecured, installment loans to consumers in the US.
- Klarna: Prime, unsecured, bullet loans to consumer at point of purchase in Europe.
- PawnGo: Subprime, secured, bullet loans to consumers in the US.
- Billfloat: Subprime, unsecured bullet loans for bill pay to consumer in the US.
- Progreso Financiero: Subprime, unsecured, installment loans to Hispanic consumers in the US.
- Capital Access Network: Subprime, secured, merchant cash advance to small businesses in the US
- Kabbage: Subprime, unsecured, installment loans for inventory purchases to eBay and Paypal merchants in the US
- OnDeck: Subprime, unsecured, installment loans to small businesses in the US.
The key variables here are credit quality [subprime, near prime, prime, superprime, new to credit], security [secured, unsecured], loan type [bullet, installment, line of credit], loan use [bill pay, inventory purchase, general] customer type [consumer, small business, large business] and geography.
These few companies hit just a few of the multitude of business models possible. They are just scratching the surface. I think we’ll see a lot more innovation in this space over the next few years.
Because of the capital requirements in making loans, these companies will need backing by large venture capital firms. A few hundred thousand dollars in seed funding will not be enough to prove or disprove whether a new underwriting model works. This is why I’m excited to meet founding teams with the technical chops and domain expertise who are interested in building disruptive lending companies – I want to make many investments in this arena.
Addendum – the same opportunity exists to apply big-data machine-learning to improve underwriting in insurance, as I have noted before.