Recent thoughts on cashflow underwriting

With Plaid stepping into lending now, the BaaS platforms offering "lending as a service" modules, cashflow underwriting "officially taking off", are we confidently saying this good old business model aka lending is the only one that makes sense in finance? 

With all the zero CAC opportunities and all the data that is available, lending seems like a "no brainer".  Consumer lending specifically is something I want to dive in today.

Having more players being the arbiters of consumer financial state and affordability should technically be great for consumers. This is an exciting future.

However the quality of the underlying data is key for having the right attributes that will then flow into the risk models and enable decisions. 

Contextually rich financial transaction data is something we were promised back in 2014. Ten years in, and we are still at it. The key problem is that in order to analyze transactions with reliable accuracy at scale, you need systems that have an understanding of the world or can change context very quickly. 

Depending on the person transacting and their overall financial situation and preferences, these are millions of unique contexts applied to every transaction and futhermore to every accountholder history. 

Then you have the merchant problem. There is a tail of merchants that does not live in any database and is not available for a lookup. Having a detailed view of merchants means you have a granular understanding of transactions, hence a granular view of consumer spend histories and their ability to re-pay a loan. 

How do you do this when there are thousands of new merchants popping up monthly ?

Current problems and solutions

At Ntropy we have been focused on solving these problems from the ground up.

First, let us look at why it is so hard and expensive to understand consumer financial behavior when all of this data already lives in your bank and wallet?  

1. Raw data is siloed and of low quality. 

Although aggregators do offer their own version of cleaning and standardisation, if you use a couple of different aggregators, which everyone does, you still get a messy output to build on top of.

2. Depending on the customer, hierarchies of spend labels need to be flexible, yet they are not. One person's Paypal transfer is paying a friend for dinner, for another person it is income they are getting from babysitting. 

3. Capturing re-payment histories from bank data only is possible, but tricky. To do this well your categorization model needs to be excellent and you need to make sure you are not missing a data source, aka a bank account. Being able to predict this is very important. Most aggregators and CRA-s will combine cash scores with repayment data in order to get here. This adds more points of friction, of course.

4. Using state of the art modelling is tricky. Billions of weights do not explain why someone was denied a loan. 

What we do

We do not build cashflow scores at Ntropy. This is not what we specialize in. We are not a CRA. We build models to analyze data. This is a form of a financial brain that will ingest and process billions and billions of heterogenous data points and turn them into an accurate and standardized view of a consumer or a business. This will reassure you that the score that you are building is the most performant and reliant one. Here is why:

1. We can give you a standard view on top of all the sources, regardless of the aggregator you are working with. This removes the need for manual mappings. 

2. We allow end to end customization, where you users can be the ultimate labelers of their own transactions, providing additional context to where and how they have spent money. This is all fed back into our models that can understand the user at the right point in time with maximal accuracy. Not just one user, but millions of them.

3. You can capture loan repayments and loan disbursements from our categorization endpoint and cross check and combine with CRA data for increased approvals.

4. Billions of weights is an issue indeed. We are working on solutions regarding explainability. 

One of them is using Ntropy's transaction cache vs calling our models. With Ntropy's transaction cache and growing merchant database the inputs into your decision engine will be the outputs of a very powerful lookup vs running an actual large model, with the performance of the latter. 

You should not do this in-house

We hear from you very often that you are working on your own solution. You should not. Build something your customers need with the best data at the back of it. This is why even aggregators are starting to work with us globally. 

Super excited for this space and for what the teams at Plaid, Alloy, Nova Credit of course, Prism Data and others are building. 

I wrote an oped in American banker recently featuring my personal experience with credit scoring algorithms and why I think industry wide collaboration is the only path towards fair access to credit and financial products that are truly empowering vs their predecessors. These were digitized loan sharks adding some very shiny marketing to it all, backed by real consumer and VC dollars.