Separating the signal from the noise in mortgage technology

As someone who started his career by designing and developing expert systems (fraud early warning, behavioral scoring etc.,), I've always been keenly interested in the science of decision making. Daniel Kahneman of course is the Nobel laureate in this field! I've been enjoying reading his latest book where he makes a compelling case for cutting out the "Noise" in human judgment and decision making process. A review of his book is here and I highly recommend it to anyone who is interested in this topic!

As I was reflecting on the theme of the book- the bias vs noise phenomenon and the techniques to make better judgments in everything from business to criminal justice reform (he makes excellent points about this), I started wondering about how this applies to our own world of mortgages (and technology in specific).

With the flood of technology solutions out in the market, there is too much noise! Lenders are constantly bombarded with solutions containing sound-bytes like AI, Intelligent Automation, Blockchain etc., It is important for lenders to separate the signal from noise. These buzzwords are great for valuation and to get the creative juices of marketers flowing, but lenders should deconstruct the mystery behind these phrases and find out what it REALLY means for their businesses.

• AI - Every other solution touts to have an "AI" component. This is a very deep topic and what is swept under the broader term of AI are really machine learning (ML) algorithms. The term "Machine learning", inherently implies that there is a "learning" component to the decision model upon which any solution is built. In order for a decision model to work effectively, the learning component of the model must have the ability to "look back" into a massive dataset to identify patterns, match with the incoming data from transaction systems, identify if there is a pattern and then predict the outcome if there is no defined precedent. In the world of transaction fraud management (going back to my roots), say a suspicious transaction is incoming, the model runs real-time and predicts if it could be a fraud or not if there is a matching pattern against the transaction dataset history. Amidst all the marketing hype, let's think hard about where this could be applied in the mortgage industry:

1. The most prevalent usage of machine learning algorithms is to recognize documents in OCR and document classification systems. However, even here, the promise is that these classification systems can recognize documents as they are thrown at them. This is farthest from the truth. Many of the doc classification systems do not have the capability to recognize documents previously unclassified and make a determination during run-time. Every new document needs to be fed into a model, trained with a dataset of examples and then rolled into production. This takes time and anything you hear to the contrary needs to be evaluated and tested critically with actual examples, rather than powerpoint ware!
2. The most often quoted use case for AI is that it is used in decision making. This is once again a creative spin. A majority of most lender's originations are conforming to the GSE and investor guidelines. So the lender is essentially following a set of defined, binary rules to originate a loan according to the guidelines. When you hear "AI-powered" decision management systems, where is the AI in this? Is it in credit, income, asset or collateral evaluation where lenders deal with uncertain data that AI will magically eliminate and increase your originations? Under the current structure of credit evaluation, if a borrower does not meet the underwriting criteria, which lender is using alternate credit scoring models that dynamically assess a borrower's ability to pay based on factors like utility bills and cell phone bill payments? Which investor today is accepting these as surrogate credit criteria? When these are relaxed and lenders are allowed to use alternate credit scoring models as a surrogate, then there is some merit to it (not to mention the fact that alternate credit scoring variables will also be defined by investors, so the model will still not deal with uncertainty!). But when a lender is originating to conform to a set of very clear, prescribed guidelines with a finite set of data points to evaluate those against, there is no AI involved in it!
3. AI - or machine learning is used quite effectively in portfolio modeling to calculate loss reserves. The models take into account historical performance of the portfolio against a set of attributes, apply it on the current portfolio and determine which accounts are likely to slip into delinquency. This is then used to create loss reserve provisioning at a portfolio level. This has been around since the time loss modeling was invented, but with the maturity of technology to harness troves of data, this has gotten better! But this is NOT originations!
4. Customer prospecting - The ability to mash up data from lead generation/aggregation services vs a set campaign criteria to identify a set of prospects has been refined using ML algorithms since the days of Credit Card Marketing by the likes of MBNA (dating myself here!) and Capital One's IBS (Information Based Strategy) became mainstream. This is a mature space in general and has the potential to improve lead conversion even in mortgages if applied correctly.

• Blockchain - We see this quoted very frequently these days. Blockchain, the technology underlying the cryptocurrencies, is indeed a very secure and robust technology to establish provenance and chain of custody without the risk of tampering. It is great for companies to be "Blockchain ready" for the future. However, a solution having a "Blockchain database" on its own today without an ecosystem and a supply chain of systems connecting into and out of it also built on blockchain is of little value.

Imagine a perfectly digital world where all 3300 counties in the US have digitized their title records, secured provenance with a blockchain enabled database, all verifications are secured by blockchain, the LOS information is secure with blockchain, pricing is locked with a blockchain and the mortgage is registered and transferred using blockchain.

THAT, is the end state where an ecosystem of connected platforms hum together in a blockchain seamlessly. Technology platforms need to evolve to support that ecosystem because that day is not too far (hopefully!), but in the current day, lenders have to take that with a grain of salt!

Much of the security that current LOS and mortgage technology demands can be met with existing data security frameworks. So lenders have to question if having a blockchain component in any of their systems causes any unintended consequences (performance delays?). While this technology is still maturing and use cases are still being figured out for the mortgage industry to be applied in a large scale, much due diligence needs to be done with the premium that is expected to be paid simply because of the presence of a blockchain component!

• Intelligent Automation vs RPA

There is often a very creative spin slapped on Robotic Process Automation (RPA) as Intelligent Automation. The fact is, there is nothing "Intelligent" in Robotic Process Automation. RPA emulates repeatable, rules-based human interactions on a computer and mimics those for execution by software programs. This space has matured greatly and smart lenders are seeing through the marketing spin that there is nothing intelligent in RPA.

RPA has a place where native LOS, servicing or systems of record do not have the capability to drive straight-through automation. We at Indecomm have our own set of pre-programmed, out-of-the-box bots for the mortgage industry called BotGenius(R) where we have automated hundreds of thousands of minutes worth of transactions previously executed by humans. One of the most common tasks that come up for automation is flood review. It's not a surprise that many lenders do not even receive XML payloads or API from their flood provider to determine digitally if a property is in a flood zone or not. Someone needs to visually look at FEMA's Standard Flood Hazard Determination Form and see if Box B.4 has a flood zone indicator in it for EVERY SINGLE LOAN!

How hard is it to deliver this data digitally and for LOS systems to natively automate this? But since both the flood provider and LOS platforms don't have this native capability, humans are doing this very basic task when they could be using their judgment to underwrite loans or clear conditions! We have automated tons of such tasks using RPA. So RPA has a place to automate such routine and rules-based tasks that unfortunately cannot be automated otherwise under the current technology solutions. But it is not intelligent by any means!

There is a place for decision management systems (making underwriting recommendations) and there is a place for RPA. One is not going to eliminate the other and any message you hear to the contrary is simply not true.

In summary, as with everything else, lenders are urged to apply context to any marketing material or spin they hear and make an informed decision about the relevance of those marketing buzzwords to the solution, pricing and ultimately their experience!

https://www.washingtonpost.com/outlook/how-to-turn-down-the-noise-that-mars-our-decision-making/2021/05/19/758be210-b370-11eb-9059-d8176b9e3798_story.html

Image attribution: Hand-drawn charts by?Mike Wolfe ?on?NoLongerSet