How To Drink From the Fire Hose of Data

Big Data is all very well, but when it comes to automated underwriting for merchant funding, the key is to know which data streams to include and which to ignore, says David Rubin.

We believe that not every Internet itch that a merchant scratches is germane to the ability or propensity to pay back an advance.

As we continue to pull ourselves out of the financial crisis, discussion about the use of automated intelligence to underwrite the merchant-funding process is again surfacing. This time around, however, there’s a new factor playing a key role in lending decisions: Big Data.

The events that led up to the financial crash are now well-documented. So, are we seeing 1995 to 2007 all over again? Well, yes and no.

With the rapid growth of alternative funding and the tech-aided speed at which advances are being written, we could well have a replay of 2007 in the making. Big-Data feeds play a starring role, adding hundreds or thousands more criteria to be crunched in the decision-making process, even as these feeds accelerate the underwriting procedure. So yes, there’s more automation, less human discernment.

But that’s not a bad thing in itself. There are some forms of behavioral data, for example, that supply multiple feeds around payment. These can be an important indicator of an applicant’s likelihood to pay back a new payment obligation. Think of it as a proxy for the character of the merchant.

Two thousand data points drawn from payment of utility bills, social-media posts, Quickbooks feeds, cash flow, and banking data could surely provide more predictive information than the 10 to 100 data points that go into a FICO score.

Technology’s Role

Douglas Merrill, former chief information officer at Google Inc. and founder of ZestFinance, says all data is credit data, whether it’s business transactions, emails, photos, surveillance videos, Web traffic, activity logs stored in giant structured databases, or unstructured data posted on blogs and social networks.

But we believe that it’s better to focus on the most relevant data, albeit available from “big” sources and derived through automated means. Far be it from me to disagree with Merrill, but we shouldn’t be lulled into the belief that making an underwriting decision based on any and all data available won’t come back to bite us big time. Some of it will sway the automated underwriting decision toward a false positive, and some will push toward a false negative.

The main complaint we’ve heard about this method is that so many data points are inaccurate from the outset, so what you wind up with is “garbage in, garbage out.” That’s when you don’t want technology to play a bigger role in making funding decisions. The emerging consumer-reporting model looks at an increasingly wide variety of data points, which may take in, but not be dependent on, FICO scores, but might also include an applicant’s Internet searches and social-media feeds. This data is combined and analyzed to arrive at a business’s credit score without requiring external credit-bureau information.

There have been some reports on how algorithmic lending can be potentially discriminatory. But a study from the National Consumer Law Center concluded, for the most part, that it was bad data that was at fault, rather than the methodology itself. We also believe that not every Internet itch that a merchant scratches is germane to the ability or propensity to pay back an advance.

Most small businesses touch a massive number of data points. We’re advocates of Big Data, but we also have a strong sense of what’s meaningful to include and what’s not. The goal is to use technology to replicate the judicious weighing processes that good underwriters have traditionally exercised.

The most innovative systems in the industry are tapping into data sources that would never previously have been considered, from Federal data to shipping volumes to social-media feeds and other publicly available information. Combining this with feeds from QuickBooks and other bookkeeping systems helps make the determination between good risks and bad risks.

So, while our ideal may be a five-minute paperless decision, the role of our technology is to rapidly crunch masses of data in a way that replicates the stringency of a paper application executed by human underwriters.

Computers do this more reliably than humans, with more consistency and less chance that human bias can play a role. Yes, human beings will build the models for the algorithm, and decide what are the most relevant and indicative data points to include in it. But we have to presume that no biases were built into the algorithm.

The Real Challenge

Some of the new breed of funders are deploying technology to more efficiently evaluate business fundamentals and to be more predictive when gauging the creditworthiness of a small business. They are using automated underwriting to help small businesses expedite the process, reducing costs and making them better able to serve Main Street businesses.

While many have developed very sophisticated algorithms on the quantitative side, there’s a real challenge when it comes to the qualitative side. With masses of unstructured data that leaves so much to interpretation, there’s much room for misinterpretation. Machine understanding, versus compilation, sorting, and reading, still has a long way to go. Hence the false positives and negatives.

But we’re also looking at criteria to help us develop a picture of prior behavioral patterns and predict future ones. These get us closer to an understanding of the borrower’s character. The more robust we can make it, the more businesses are opened up to get funding.

Bottom line: In using Big-Data technology to evaluate risk and fund today, as opposed to tomorrow or next week, what you choose to keep out of the mix can be just as important as what you put in.

—David Rubin is the founder and chief executive of eProdigy Financial LLC, New York, N.Y. Reach him at ceo@e-Prodigy.com.