How complexity in model offerings can actually be a smoke screen for weak technology

The blockbuster movie, Everything Everywhere All at Once offers a look at the world from the perspective of how every decision creates a multitude of universes following the paths of the alternate choices that could have been made.

Now, that’s a pretty complex narrative to follow, and as the film shows us, all of these “what ifs” lead to confrontations, confusion, and an over-complication of what could have been solved by simply speaking to another person.

The technology for lenders today — whether you’re using it or looking into it — isn’t one-size-fits-all, but it also doesn’t have to be over-complex. If we get caught up in the complexity of a model when it comes to AI-automated underwriting, we might miss finding the simpler, and more effective, solution to the problem we’re facing.

You don’t always need to jump through a thousand hoops (or to compare with Everything Everywhere All at Once — a thousand parallel universes) to achieve better lending strategies and processes. And if your technology partner is only offering you complex solutions, the question I’d ask is if that technology company is offering you a complex solution because there is no simple answer or if they cannot offer you simplicity because their technology isn’t up to snuff.

‍

‍Complexity doesn’t directly correspond with ‘better’‍

Albert Einstein once astutely observed, “Everything should be made as simple as possible, but not simpler.”
‍
This mantra deeply resonates with Zest AI’s approach to building powerful machine learning credit models for our lending partners. We strive to create models that capture every loan applicant’s complex, nuanced set of data and elevate fairer outcomes across borrowers. Yet, we know it’s important to avoid superfluous complexity that could make our models cost-prohibitive, unstable, noncompliant, or dysfunctional.

Injecting complexity everywhere, all at once, is easy, and it might even look like the right thing to do. After all, lending is a complicated process — so one might think that the technology you need has to match that complexity. Actually, let’s take a second to walk through what that might look like: all a lender and that technology partner would have to do is take as much data, create as many modeling variables as possible, and put them into the most complex machine learning algorithm imaginable.

Such a model may yield statistical performance, but it will undoubtedly increase operating costs, require additional regulatory scrutiny, and add to your operational complexity.

‍

‘Better’ takes time, tenacity, and fine-tuning‍

Over Zest AI’s fifteen-year history, we have perfected the use of AI in credit underwriting. We have built some of the largest ensembled models utilizing data from credit bureaus, alternative sources, and core banking information shared by our clients. We have also developed highly accurate, tailored models with just a single source of data. We’ve taken the time to understand these forms of credit modeling better, fine-tuned them to create optimal outcomes for our clients and their borrowers and have the tenacity to keep improving our methods to deliver the best available outcomes — complex or simple.

Our commitment at Zest AI is to meticulously tailor our models to meet our client’s specific needs and help them achieve their strategic business goals and objectives.

We start by sitting down with our client and discerning the model’s intent, the demographic it will serve, and the desired business results our client wants to achieve. This understanding allows us to efficiently determine which available data sources offer the best return on investment. We do this by considering the improved performance against additional data costs — both in tangible and intangible terms.

Ultimately, when we’re digging into our clients’ business strategies, what we want to achieve is the simplest answer to whatever-level complexity of the problem they face. We don’t tack on complicated algorithms as a handwave solution to show how remarkable our technology is — we think the most impressive solution is the one that is straightforward in answering a need.

‍

Opportunity costs and what opportunities cost

As a general sentiment, our clients aspire to automate lending for all their borrowers. That’s a huge opportunity to speed up decisions and get that hard-won loyalty from a borrower. The way we see it is, the simpler it is to get from point A to point B, the more value we bring to our clients.

However, the stage of the automation journey that a client is in can vary, thereby influencing their need for additional data sources. This is where it becomes important to determine how much complexity is actually needed to meet that automation goal.

If a lender is only automating the top 25 percent of applicants, the economic benefit of better rank ordering the borrowers in the bottom 50 percent may be insignificant. On the other hand, lenders automating over 80 percent of decisions and open to lending across the spectrum might find alternative data immensely beneficial once data costs are factored in.

The addition of a new data source in a model comes with certain costs:

• Direct data costs: each new source incurs a cost per data pull, which can quickly accumulate for lenders with high application volumes and low funding rates.
• Technical costs: integrating new data sources requires developing appropriate integrations.
• Complexity costs: managing multiple new vendors can strain internal resources, and using different data sources from various vendors may lead to model instability due to data inconsistency or volatility.

When deciding what data to incorporate into our client’s machine learning model, we always consider these costs and potential benefits. We are transparent and conduct a comprehensive proof of concept and ROI analysis that allows us and our clients to make judicious business decisions on which data to utilize in their model. We also perform extensive fair lending testing on the data sources to ensure that our credit underwriting models foster inclusivity and fairness for all borrowers.

We generally advise lenders to begin with simple models and gradually introduce layers of optimization. As they grow more comfortable with machine learning and high levels of automation, alternative data sources can be evaluated and incorporated at an appropriate time. This approach ensures lenders reap maximum benefits from their investment while avoiding unnecessary complexity.

One thing I walked away with after watching Everything Everywhere All at Once was the idea that sometimes it’s not the complex, verse-hopping technology that saves the world — it’s the simplicity in valuing one another and seeing through the complexity and into humanity. At Zest AI, we do have that technology, and we’re happy to help our clients make use of it, but we first seek to understand which is what makes us a technology partner worth our weight in gold.

‍

Want to read more about how to evaluate if a technology partner is the right fit for your organization? Read on in this blog here.

____________________________________

Adam Kleinman — Head of Strategy and Product

Adam spends most of his days in conversation with Zest AI’s clients, finding pain points and bringing innovative solutions to the table to push the lending industry towards a more efficient and inclusive future. He’s built his career within the financial services sector and believes that AI can be a powerful tool for building a strong economic ecosystem.