Prototyping the Learning System

Fundbox's product strategy

For the management team, access to data was critical.

The product vision at Fundbox was built atop two key fundamental philosophies, as Fuloria explained:

“First, there was a market need for working capital that we noticed by talking to customers and looking at the market data. Applying for a loan at the bank is a very tedious process that takes a lot of time and often ends in people getting rejected. Second, we believed that we could take a broad technology approach to solve these small businesses’ issues thanks to increased availability of data and the creation of machine learning tools that could make the most of it.”

Early in the company's journey, the cofounders had noticed a trend among small business owners. In particular, they found that many were becoming increasingly comfortable with putting their data in cloud-based accounting or invoicing platforms like QuickBooks, Xero, FreshBooks, and others.

Fuloria commented: “With that data in the cloud, we knew that it would probably be possible to access that data via an API, which we could then use to analyze customer behavior in real time.”

When it came to designing a learning system that could help the company continue to learn more about its customers and adapt to their needs, Fuloria and his team were very cautious about taking an approach to underwriting revolving credit for small businesses that would be similar to that of incumbents, as he explained:

"Using credit in the traditional way was not a focus of ours in the beginning. If we were to have taken a conventional approach, we could have hired a chief risk officer from a bank, who would have immediately recommended that we encode the best practices in the world of risk management into software and have us start by focusing on customers that were the safest from a traditional credit perspective. However, as it turns out, there are plenty of small businesses that are creditworthy, even if the owner may not have a spectacular personal credit score.

If you start with a very traditional approach to the solution, or if you have that in your company’s DNA very early on, then you may get to a different outcome altogether."

Built to Learn

From the beginning, Fundbox’s management team set the company up in such a way to learn as quickly as possible about customers’ pain points and needs. The team started by building a product that focused on a revolving line of credit for small business owners. Revolvers allowed for the Fundbox underwriting system, which focused on leveraging machine learning to predict customer’s behaviors better, to gather as much data as quickly as possible, as Fuloria explained:

A straightforward way to think about data is to use a matrix where the rows are the number of observations and the columns are attributes associated with those observations. For example, when I was working on the ad systems at Google and Facebook, each time we showed an ad to a user we counted that as an observation. After the user would see the ad, they could click on it or not, and if they clicked on it, they could potentially go on to convert—whether that meant buying a product on Amazon or completing some offline behavior, for example.

In these systems, it was simple for us to learn thanks to the sheer volume of the number of data points. If you’re serving billions of ad impressions every day, this is in the billions. This made it easy to run an A/B test and learn, within hours, minutes, or days, whether one advertising format worked better than another. All of this data also helps train machine learning models to predict the likelihood of a click very quickly.

Now, when you switch from the world of advertising to the world of financial service, and in this case, lending, all of the sudden you’re not receiving billions of data points. Even today, Fundbox has a little over 100,000 connected customers. Each observation isn’t checking to see if a website visitor clicked on an ad, but is instead a matter of extending credit to a customer. So now, we’re talking about tens of thousands or hundreds of thousands of observations, over the course of a year or more, as opposed to billions in a much shorter timespan.

In addition to fewer data points over a longer time horizon, each observation carries more risk. If I’m extending credit to a user, the positive result is that they pay us back. However, the adverse effect is that they are unable to pay, and Fundbox loses its principal capital. So not only do we have an order of magnitude fewer observations, but the cost of an observation is much higher.

So the challenge is—if you genuinely want to build a systemized and automated approach to making predictions about lending to a business, how do you develop and train your models with relatively little data?

(See Exhibit 2 for a comparison of the advertising and lending models)

**Exhibit 2: Comparison of advertising observations and lending observations**

In addition to allowing business owners to link their bank accounts to Fundbox, Fuloria and his team focused on gaining access to user data by allowing potential customers to link their company’s online accounting software to Fundbox, which allowed the company to underwrite based on the business transactions included in the accounting software.