Where’s the disruption at?

May 23, 2018

Over the last decade we’ve all heard a continuous drumbeat of reporting on how Amazon is destroying traditional retailers. Same for the battering print and television companies are taking under the two-pronged assault of digital content absconding with eyeballs and digital advertising delivering a superior product for marketers. But the prognostications of doom for consumer banks have basically stopped. How come?

After working inside banks and alongside fintech lenders and investors here are some of my thoughts on the matter. Like most things involving consumer credit, the factors are interrelated and compounding:

There ain’t no such thing as a free lunchAcquisition Costs

A venture capital partner recently told me the first question they asked after consumer fintech pitches was whether the company had a way to create a proprietary and low cost customer acquisition channel. If not, hard pass. A sensible and common stance, especially if you want your cash directed toward engineers building the product of the future not toward marketing or fending off some copycat competitor. But what happens if the proprietary and low cost channel is only a fraction of the overall market, what’s a founder to do?

Enter the reliable fly-wheel standby:

Product can mean a lot of things here but for consumer and small business fintech it means a combination of customer experience and financial value proposition on the consumer side and model driven predictions of customer performance on the companies side. The theory goes that if consumers think you have a great product (experience + value), people will talk about it, creating viral marketing ‘network effects’ on your behalf that will lead to zero CTA users who in turn will lead to more users. But here’s the thing, people don’t talk about financial choices with their friends as much as they talk about a new streaming service or ride hailing app. They also don’t make banking decisions on a daily basis (I’m not counting making a purchase with a debit/credit card), further weakening any network effects.

Aside from not driving sufficient volume, network effects also have another issue. They do not allow for the same degree of control and refinement of customer selection that paid marketing does. All users aren’t created equal in that they won’t have the same risk and revenue performance. Paying for good performance is better than getting bad performance for free.

Smokey this isn’t Nam eCommerce, there are rulesRegulation

Much of modern data science is performed using open-source tools but the adoption and improvement of those tools was and continues to be hugely influenced by companies engaged in some form of ad sales or ecommerce.

If I’m a data scientist in one of those shops my primary goal is to create conversion lift which in turn creates incremental revenue. At my disposal is basically any data on the customer I can collect from their interaction on my sites/apps or that I can purchase from a data broker. Gender? Yep. Exact address? Sure thing. Basket data from my last trip to the drug store? If I’ve got good data coverage from a broker, why not? The primary limitation on what was usable was what was available and then how quickly you needed the model to run. However, if that model is being used to influence a consumer banking offer you have to comply with the Equal Credit Opportunity Act (ECOA). Not only will this severely restrict the data set available but it also requires any negative decision come with an explanation of why that decision was made. Difficult to do in a 500 variable, non-linear model.

Because much of modern data science is taught on the job, most candidates for fintech data science jobs come from adtech, ecommerce or directly from academia. As a result I’ve seen plenty of risk model lift charts demonstrating the prowess of a new “proprietary algorithm” that was built on credit bureau and application data (data any lender would have), only to find off limits variables and features. For example, a model using 9-digit zip code and one that categorized applicant first and last names. The models could geographically redline to almost household precision and picked up a signal from name that inferred the racial background of the applicant. They both predicted risk, but they also predicted race and gender which meant they were non-compliant with ECOA. In other words, if a new company has a better model than incumbents built on less relevant sample, it probably means they are knowingly or unknowingly cheating.

The waiting gameDelayed Outcomes

Beside regulation, another huge difference between adtech/ecommerce algorithms and consumer finance algorithms is that the length of time between prediction and actual outcome. If a consumer sees a display ad there some finite window of time for any subsequent purchase of the product to be attributed back to the ad. It could be days, weeks or maybe a couple months. When you only have a week or two to wait for new training data to arrive, the amount of treatments you can test and iteratively improve upon in a year is fairly large and probably constrained by data scientist resources or sample sizes more than data availability. For financial products there is a lag between a modeled decision and an actual outcome that is measured in months or years.

I think this is largely understood for credit risk but solving the problem using statistical methods requires a lot of time and a lot of sample. A popular short cut has been some combination of very conservative haircuts and loans to family and Stanford Business School classmates (early adopters!) to demonstrate risk management chops. Expansions beyond that rarefied 0% loss rate cohort require eating losses for a risk model to evolve. But sometimes the revenue side can be even more tricky. Many credit cards, particularly targeted at low risk customers with plenty of choice, rely on performance in 5th+ year of a customer relationship to drive the overall economics of the product, subsidizing early losses and marketing costs.

Because credit losses and revenue are unknown without historical data, testing is critical. But it’s expensive and requires a lot of time to mature. Pile on expensive credit facilities for funding and unit economics become really challenged.

Means and mettleConclusions

But there are reasons to believe that fundamental transformation is still on the way.

The incumbents are entrenched with consumers on the stickiest product (checking), have a huge advantage in performance data history (multi-product and through-the-cycle), cheap funding and meaningful marketing budgets. But they have organizational, technological and legal barriers to leveraging that data advantage and large commercial banking divisions that prevent a laser focus on consumer products and often demand concessions from their retail arms in dealing with tech companies to support investment banking relationships.

SoFi’s banking charter application and subsequent announcement of a checking account product is the best US example of how fintechs will need to mature to put meaningful pressure on incumbents. Not just to reduce funding cost, but to cement longer term customer relationships and build the multi-product customer lifetime valuations that modern data science can support.

Amazon’s partnership with JPMorgan on checking accounts feels akin to inviting the Trojan army into Troy without even the pretense of a giant wooden horse. Armed with checking account data, Amazon will be able to optimize capturing banking profit pools by intermediating customer relationships and taxing any banks down stream via revenue sharing or product acquisition bounties.

Both SoFi and Amazon understand the key is the data collected throughout a customer’s lifetime, over many interactions. The data is what provides for a better product valprop and customer experience. The nuances for financial service products is that the data takes longer to mature, can be brutally expensive to acquire and incremental business with a customer can lead to negative economic outcomes for the bank (i.e. losses, low revenue not covering OpEx).

The future will belong to evolved incumbents or de novo banks that leverage modern data science and a unified multi-product data layer to predict customer lifetime values earlier than competitors and acquire those customers, if necessary, with loss leading products like high interest checking accounts with no fees. They will then have a substantial proprietary data advantage to maximize overall customer lifetime value. Those without this capability will face negative selection, volume reduction and eventually extinction.

Jackson Barnes is a Partner at Full Spectrum Analytics, an analytics and data science consultancy specializing in financial services.