Artificial Intelligence And Mortgage Banking: Opportunities and Perils
By Rebecca Walzak
March 27, 2018
Rebecca B. Walzak, a 40-year veteran in consumer lending, is recognized as a leading advocate of Operational Risk Management and a supporter of technology associated with managing risk. She is president of rjb Walzak Consulting and is a partner in the MarBecca Group. She is responsible for providing consulting services to clients in Quality Control and Regulatory Compliance as well as Operational and Enterprise Risk Management. She has also provided expert witness reporting and testimony in numerous litigation matters.
Walzak has served as Chair of the MBA Quality Assurance Committee and has been a frequent lecturer at MBA and brokers' conferences. In addition, she continues to author numerous articles about the industry and the need for operational controls, including the advancement of technology and its impact on risk management. She currently serves on the Board of Directors of the MBA of the Palm Beaches in Boca Raton, Fla.
Artificial Intelligence is coming!
These words conjure up both hope and fear in many people. Hope for mortgage lenders that this is the answer to productivity costs, over-burdened staff and repurchase risk. And fear that our jobs will disappear because of it. However, most people, including those in our industry, just don't understand what AI is and what it isn't; what it will do for them or to them.
But why? Is it just the lack of understanding about what it really is? Or is it an aggrandized fear stemming from numerous late-night horror flicks. Could it be the hope that in the future, diseases, poverty and hunger will become a thing of the past or as is feared by some, that robots with greater intelligence will rule the world. Neither of these views is unique. Well known individuals such as Stephen Hawking and Bill Gates see it as the end of the human race as we know it. Yet futurists such as Ray Kurzweil, the author of "The Singularity Is Near" and "How to Create a Mind", see it as the opportunity that will propel humankind's best morals, values and humanity to the highest levels.
"Singularity" by the way, is the point at which machine thinking is equivalent to human thinking as established by the Turning test which compares human intellect with a machine's. This, according to Kurzweil, will possibly occur as early as 2029 and most definitely in the early 2030s. Before that however, as early as the mid-2020s, most individuals working on or in the field of artificial intelligence believe there will be significant advancements in the use of AI and robotics.
Even today there are precursors of what this technology will bring. Take for example the technology that provides for 3-D printing. While we recognize that this is available today, many people think only of what "fun" things we can do with it. However, according to Kurzweil, expanded development of 3-D printing will allow us to print out much of what we need to survive--for example, the clothing we wear and the shelter we need. With hydroponic plants for fruits and vegetables we will be able to create food inexpensively using 3-D vertical agriculture and in-vitro cloning of muscle tissue that will provide meat relatively inexpensively. The fact that these programs will be available on open-source technology creates an unlimited opportunity for its use and levels the playing field for those individuals not involved in its development.
The impact of these advancements on what we do for "work" is also under discussion. The Pew Research Center recently published a paper by Aaron Smith and Janna Anderson, AI, Robotics and the Future of Jobs. To develop the thoughts and opinions of recognized experts in the field they sent out a survey asking these experts to provide their thoughts on the subject. The results offered by them are mixed although almost evenly split between "good" (48%) and "bad" (52%).
Smith and Anderson found despite the very different outlooks on this technology, experts agree that robotics and AI will permeate a wide segment of daily life by 2025. Those who thought positively about AI were inclined to recognize that advances in technology have historically been a net creator of jobs even as it eliminates or displaces some types of work. The new jobs created by technology will result in the invention of new types of work that take advantage of the technology but will also create jobs focused on human capabilities that are unique, such as ethic-based decisions.
We are already seeing this occur. For example, who would have thought that "social-media marketing" or "search engine optimization" would be such significant job titles as they are today. Probably the most agreed-upon positive effect will be the ability of this technology to free us of the day to day work that is just plain drudgery. Just think of Roomba that now does the vacuuming whenever it is told to. You'll never get that level of obedience from your kids. Overall, it was their belief that "work" for humans will become more positive and socially beneficial.
Those experts negatively inclined toward the advancement of AI and robotics had a different view on the subject. Their overwhelming expertise, while recognizing the net creation of jobs from technology, envisions that the disruption in the workforce will go far beyond simple daily drudgery work. It will in fact interfere with both blue-collar and white-collar jobs. While it is a given in their minds that highly skilled workers will succeed, others displaced by the technology will end up with lower paying service industry jobs or face permanent unemployment. In either scenario these concepts, beliefs and experience are the strongest indicator that everyone, whether in the working world or not, needs to understand much more about this technology and how it will affect them.
What is AI?
In a paper by Patrick Hall, Wen Phan and Katie Whitson, The Evolution of Analytics, the authors provide a detailed look at numerous options and ideas for the analytic development and functioning of artificial Intelligence. However, for those not as technology savvy as these authors, the most simplistic explanation is that artificial intelligent means training and programming a machine to give it cognition or things that are associated with learning and problem solving. Similar definitions describe it as "the ability of a computer or a machine to think and learn." Underneath the umbrella of AI there are distinct areas, two of which are Neural Networks and Expert Systems. Mortgage lenders are most familiar with Expert Systems but are becoming more familiar with Neural Networks as the technology advances.
What's the difference between the two? Expert Systems are a method of automated reasoning and include three items: (1) a set of facts, rules and principles. (2) a set of data that is a question to be solved by these facts, rules and principles and (3) a user interface which will filter the data to conform to the rules in the system. We are most familiar with these systems from the automated underwriting engines we use daily.
In these Expert Systems, the rules are the underwriting standards that are found in a company's credit policy. The data come from the loan application and the underwriter is the user interface. Beside these rules-based systems, other expert system types include Case-Based systems and Decision-Tree systems. Case-based systems import an unlimited number of cases to the system. When a specific issue arises, the specifications for that issue are input into the system and is matched to a previous case to arrive at the best answer. Case-based systems are most frequently used in areas such as the diagnosis of medical conditions but are also being used in the development of self-driving cars.
Decision trees are based on inductive reasoning with the outcome expected to be found at the end of the path when it indicates a solution to the problem. This type of expert system is also used in some of the AUS programs today. More encompassing AI programs are now known as RPAs or Robotic Processing Automation because they mimic user activities, process structured data, are highly determinate and involve the existing labor force for assistance.
Neural Networks involve the "teaching" of a system's neurons to respond as a human brain does. A neuron is the basic working unit of the brain, a specialized cell designed to transmit information to other nerve cells, muscles, or glands. Just as a human brain contains millions of neurons that send information received to other neurons, AI programs create neurons that send information to other neurons in the system until a pattern of information is established and an answer can be given, or an action taken.
Neural Networks are the unpinning of Machine Learning. These AI programs are based on the construction of algorithms which can "learn" and make predictions on the data they receive; in other words, neurons. These algorithms include such functions as regression analysis, decision trees, and support vector machines. The system then identifies patterns in this data and learns to use it in unassisted ways.
Obviously one of the major elements in these systems is that they receive lots and lots of data from across a nearly infinite network. This includes large amounts of complex and unorganized data. Complex and unorganized data is often found to be dirty, noisy or unstructured and available most anywhere. These learning models can restructure the data and make use of millions or even billions of parameters.
The "training of these machines" occurs by running this tremendous amount of data through the algorithmic model until it finds enough patterns to be able to make accurate decisions on the data. It then scores the new data to make predictions. These new predictions are stored and incorporated into even more data. Currently these models are being used to make recommendations on such actions to be taken as fraud detection, patterns and image recognition.
Companies that offer specific marketing leads or companies that want to focus on just once specific segment of their market can use these machines to find individuals that fit the profile they are trying to attract and then send emails or mail flyers to perspective clients. Some of the best values seen for banking entities include the ability to balance risk far better than using just humans, the ability to detect fraud and to market new products. Since the industry has recently begun to incorporate a manufacturing focus on production and servicing, the value added by AI for lenders is seen in its ability to detect previously undetectable defects. Potentially this would eliminate the need for back-end reviews as the risk factors in operational failures would be easily identified and corrected without having a large costly workforce of individuals whose work is a "stare and compare" methodology.
While the impact of machine learning is beginning to be felt in most industries, AI developers are not through with enabling machines to think like humans. "Deep Learning," a higher level AI, uses the approach of machine learning but, as the name suggests, go further into cognitive computing where systems perform specific human-like tasks in an intelligent way. Deep learning combines the ability to translate problems from natural language into machine logic, query a knowledge base and use machine learning algorithms to make decisions about potential solutions. "Watson," the IBM AI machine of Jeopardy fame, is probably the best-known in this area.
Of course, there is so much more to be included in this transition to an artificially intelligent world, but there are also problems and perils that all industries must address, as well as those unique to mortgage lenders.
Artificial Intelligent machines are coming so much sooner than was anticipated, and most see this as a good thing. There are however numerous problems facing implementation of this technology in any industry. Two of the most common cited are organizational and ethical challenges.
Organizational changes, while on the surface seem rather obvious to many, are not simply eliminating staff positions with new technology, but require overall process re-designs and new functional position requirements. Today we are enamored with the idea that applicants can simply enter application data into a phone and get an approval, however the long-term implications are much more significant.
Organization changes must be focused on the long-lasting impact of AI tools and can only be made when a fundamental shift in the organization's culture is in place to support these changes. For example, if a lender has a culture of allowing loan officers to refute underwriting decisions, gather more data for underwriters and push to get loans closed before the end of the month, what will happen when loan officers become the "user interface" with the technology.
In this scenario, the borrower simply supplies basic information, the AI tools gathers all additional information from a generic database of consumer information, evaluates the loan using data prediction analysis and gives a score that identifies the probability of a default, establishes the interest rate based on the true risk of default and identifies any conditions. Once this is complete it will obtain and evaluate title commitment issues, deliver the documents to the borrower who will sign electronically and then funds the loan. Known as STP or straight-through processing, this concept results in a process being completed end-to-end without human intervention.
How will the culture that centers on customer support by loan officers continue to exist In this environment? Will loan officers cease to exist? Instead of loan officers, what type of staff will this company need?
This is the dilemma that will be facing every industry in the future. While there will continue to be the need for customer service support, what other type of talent will be needed? Most of the job functions required for future work are just not there today. In "The Evolution of Analytics" the authors state that there are just not enough people with the skill sets to develop and execute analytic projects involving machine learning nor are there enough people to drive the change management that is necessary to develop a data-driven organization.
In terms of future work, there is a real shortage of deep analytic talent. This is especially true in our industry. For years, reports as simple as the results of quality control reviews have been ignored because management does not understand the relationship between a statistically valid sample and the entire population.
How then, can we expect current management to understand the relationships developed by neural networks that result in multiple reports based on this methodology and use the information in a positive manner?
And speaking of data, the future of work from an AI perspective is all about the data. Companies will need data scientists that are trained and skilled in computer science, math and domain expertise. These people will be the heart of the organization's ability to deliver the products and services that meet the value proposition. In addition to the development and training of AI technology, businesses must also be focused on enabling processes and technology that support and enable staff that are less analytically skilled. Having AI analyze and identify why loans fail does not assist a collector when discussing resolution options with a defaulted consumer.
The attempts and headway made on creating consistent reliable data is only the tip of the iceberg in this industry. In industries where AI is already used more effectively there is a data-culture and decisions made on this data is expected. Medicine is a great example. With the work done through AI, doctors are able to receive assistance in diagnosis and treatment of medical conditions. Doctors also routinely input vital data about patients and the improvements or decline in their health. These data are then used to conduct more analytics and refine the recommendations a doctor receives. These data have also been used to discover ways and means to cure diseases that just a few years ago were considered death sentences.
To reach this point organizations must have a data culture. In such an organization, data-driven decisions are required and those made on "a gut feeling" or data that are not accurate or complete will no longer be acceptable. To achieve such an organization, businesses must implement rigorous data collection techniques. This includes implementing required business processes and investing in the technology that will capitalize on AI. Having technology that gathers the same data in different formats, or using "notes" input by people defeats the ability of the technology to advance business.
Underlying this data-driven organization is the essential need to get accurate results from any AI program including expert systems. There is no doubt that the Great Recession was led by loans originated and approved through these systems. Ultimately these systems were found to be imbedded with false and inaccurate data. One only has to look at the numerous lawsuits and monetary awards to understand what false data can do.
Of course, false data or fraud, is not the only data issue. There are numerous issues including "noisy data" which contains conflicting or misleading information and "dirty data" which has missing, inconsistent and erroneous values. It is becoming more and more evident that companies that rely on data, must have an organizational function that is charged with the security and governance of data. This group is responsible to ensure that all data is secure and answer questions on how the data should be managed.
Another area of concern for AI processing involves regulatory requirements. Not the ones that involve providing disclosures or even having work completed within specified time frames, but those that involve the potential for biases in the decision-making. Companies that mine data and others that use collected data have found that the analytic patterns developed by these tools, can, without human intervention, make biased decisions. For example, if a company wants to reach out to people to offer a particular product, they may search databases to find the attributes of people who have bought similar products previously. If during this search, the database finds that most people who drive Jeeps have not purchased these products, then the listing of individuals who receive this company's offer excludes owners of Jeeps. Imagine this scenario, only the product offered is a mortgage refinance. The resulting "Fair Lending" issues are presently unknown, but could have significant complications for lenders.
There is no doubt that this technology will have a lasting impact on the way "work" is done. However, there are broader issues still to be faced. Wendell Wallach, a pioneer in the field of robot ethics, believes that this transformation has deep societal implications that we are just beginning to comprehend, let alone resolve. For example, if AI is successful in eliminating disease and other societal problems that shorten life-spans, will we live twice as long? What is the impact on society if people live to be 150 years old, or even longer? The impact on society also involves cultural and moral issues. After all, according to Wallach, we are as much a product of our culture as our biology.
One theory being presented is that of Artificial Moral Agents that augment the new reality presented by AI. These AI moral agents would be designed and trained to implement moral decisions that reflect the standards of specific cultures. Another thought is that in the creation of all AI machines, a moral agent would be imbedded so that the results have ethical sensitivity.
Who does this new moral, ethical society depend on for the development of such morality? Our morals and values come from not only internal ethical standards, but from society in general. These questions have yet to be answered. What role will personal responsibility, temperance and discipline play when any urge can be gratified at almost the same moment it is felt? With all the activities that are foreseen to be conducted by robots, what will happen when the only decision to be made is a moral one. For example, what decision will a driverless car make when faced with turning left into on-coming traffic and destroying a car with a human driver or continue going straight ahead and running over a woman and child in an intersection?
What will warfare look like? Will it be nanobot-based weapons as well as cyberweapons? How will decisions be made about when these weapons will be used and who is expendable? And what does this mean for the world as we know it with "nation-states" that interact with each other and make decisions that impact the global community. Will we ultimately eliminate these artificial boundaries and identify as members of a world based community. And who will make that decision. Wallach states that "In a democratic society the public should give at least tacit approval to the future it is creating" or ultimately, as Hawking and Gates envision, will this technology become an existential risk that threatens the survival of our civilization.
Mortgage Lending and AI
Throughout the work of the futurists reviewed for this article, they identified three stages in the acceptance of artificial intelligence. Included in these stages are, first is awe and wonder at its potential, next comes a sense of dread at the grave new dangers it produces and the finally the final stage of establishing an approach that can realize the benefits while managing the dangers. So where is the mortgage lending industry in its acceptance of AI?
It is clear from the activities going on that the industry is firmly in stage one and has been for a while. The first expert systems were introduced in the early 1990s and unconditionally accepted by the mid- 2000s when there was overwhelming volume to address. This advancement stalled between 2008 and 2015 as we recovered from the effects of the Great Recession and battled to hold our own in the deluge of new regulations that were legislated.
Recently we have begun to once again experience the "awe and wonder" as digital mortgage lenders are offering applications by phone and approvals in less than 10 minutes. This process has already become a prominent feature for most lenders. As evidenced by the numerous vendors at the national convention last fall, we are also beginning to expand into other opportunities that AI can provide. For example, Fannie Mae is working with IBM to develop and utilize its capabilities in a variety of ways. Other lenders are working on developing "bots" that can answer routine customer calls, relieving staff of this responsibility.
However, we have not experienced the return we expected. For example, recent information indicates that the digital mortgage is not returning the benefits thought to be forthcoming with its implementation. While the industry has experienced high volumes of applications on these systems, the actual close rate, as stated by Freddie Mac, is only around 14%. The reasons for this low close rate are still not clear and the benefit of more volume with less work for loan officers has not been established. For example, loan officers associated with these applications have not had to spend less amount of time with applicants. Instead in some cases it takes more time. Neither has the value of a digital mortgage approval been accepted as expected in the realtor community.
One area which was expected to deliver better results was the use of digital mortgages by the millennial population. Due to the overwhelming evidence that millennials rely on electronics for most everyday activities and resist talking on the phone or physically going into a store to buy something, appears to be true only for everyday purchases. Numerous surveys and discussions with loan officers indicate that these individuals are most often uncomfortable completing an application without the support of an "expert" to guide them through the process. One loan officer told me that most millennials are more likely to call him prior to entering information (some even field by field), wanting to be sure that they understand what is being asked.
On the other hand, baby boomers are more likely to enter what information they know without completing the entire application and can't seem to understand why they should complete the whole thing. In both of these cases, loan officers have said that they spend more time on the phone with these individuals reassuring them of what is needed or just getting the rest of the information that was needed for a decision to be reached. Another loan officer told me that he liked the digital mortgage application because he could meet with the applicant while entering the necessary information when he was with them to make sure everything was complete. This isn't much different that the on-line application that has been around for years. One thing that was also evident was that the overwhelming majority of people using digital mortgage applications did not know how to electronically include paper documentation by scanning and sending it.
My discussions with those involved in digital mortgages also found that most people were uncomfortable signing the legal documentation without understanding what it was telling them and did not have the time to read it themselves. Once again, they wanted to have someone available to answer questions at the closing. It is evident that there is much to be done before a mortgage application is a straight through process and the benefits, both monetary and efficiency are achieved.
Of those issues yet to be addressed the delay is not necessarily one of acceptance of the technology, but the significant process and change management yet to come. Despite the vision of significant reductions in staffing that AI promises, nothing appears to have changed once the digital application is input. While any number of applications are approved, people are still needed to explain the approval and answer questions about rates, fees and conditions.
If the loan is not approved, rather than just deny it, it is sent to the original loan officer or, if the applicant has not talked to a loan officer previously, a loan officer is assigned to address the problems and work with the borrower to gain an approval. This sales staff, whether internal or not, then takes on the task of obtaining documents, clarifying the information or switching the borrower to a different loan program for which they will qualify. Underwriters are still needed since the basic automated underwriting system used is most likely DU or LP. When loans fall outside of their parameters, underwriters must step in to see if the loan is acceptable. Beyond the digital delivery of closing documents, little appears to have changed in the closing and funding area. Post-closing, delivery and quality control appear to have felt no significant impact from AI at this time.
Unfortunately, these non-revenue areas are the ones with the most staff and whose primary function is simply "stare and compare" functions. These are the exact functions which AI is designed to take on but have been ignored in this industry. While defect identification is a major function of AI, its ability to detect previously undetectable defects lends itself to these back-room functions as well. It also appears that currently, efforts at change management and human functions have had very little attention as has the oversight and rigorous efforts need to control data.
A major problem for this industry is that of data. For years we have been trying to develop standardized data sets and ensure accurate reporting. Despite the efforts and success of MISMO, there are still entities that retain proprietary data fields and/or do not follow MISMO standards. This reluctance of lenders, both bank-owned and independent, to accurately collect as well as share data prevents the industry from collecting the data necessary to use AI tools effectively.
While some may believe that the work Fannie Mae is doing with IBM will demonstrate that the issues will be solved, but that is not the case. Fannie Mae is still just one body who holds data. For AI to be effective for all lenders it must have data contributions from all lenders. In addition, there is still some vital data that are not collected or utilized by the industry that negates whatever findings there are today. Operational defects that produce inaccurate data are not captured although they have been shown to have significant impact on loan performance.
In addition to collecting data, the issue of governing the data collected does not appear to be a major concern of the industry at this time. One of the major lenders with a digital mortgage program has added a function for Data Governance but its actual work and responsibilities are not yet clear. Data security has received a significant amount of attention since the requirement for data privacy has been passed. This however does not mean that security is tight as recent episodes have shown.
Progress also needs to be made on how a company using AI will be organized and managed. The job functions described earlier will need to be integrated into the technology area to maintain and enhance the intelligence upon which the organization depends. Based on research it is likely that the jobs that we have today will be gone. In their place will be positions that are primarily focused on customer service; more of which will be in the servicing arena. The people filling these positions will be held by those with extensive knowledge of both the origination and servicing functions and with the ability to use technology effectively.
In addition, there will be "ethics-based" decision makers that will deal with questions regarding the moral and legal decisions to be made. For example, if a major customer of a bank makes an application for a primary residential mortgage loan on a property that the system knows is a rental and rejects the application, what decision should be made? Should the bank go forward with the transaction and override the AI rejection and give in to the customer's demand?
What will happen to the staff, both clerical and management functions, that do not meet the requirements for these new positions? Will they be out of a job or retained for another position in the company? The 52% of those surveyed who have the most negative vision of AI are worried that there will be a significant displacement of both blue and white-collar workers, leading to greater income inequity and groups of people who are unemployable which will cause significant social unrest. The idea of mass training programs to retrain some mortgage staff does not appear to be a point of discussion.
Envisioning the Future
The industry has jumped in a big way into the idea of artificial intelligence and what it will provide. Obviously, restructuring the organization with far fewer people and the resulting reduction in risk and costs is the most significant driver of this movement. The thrill of being involved in leading-edge technology is also part of the draw.
However, the secondary market also has a lot to gain. Collecting accurate data, scoring on both credit and operational issues and analytics that give accurate risk measurements from which pricing can be based are part of the equation. This ability to effectively price the risk of loans will most likely draw more investors to mortgage-backed securities and ease the standards that drove the multiple lawsuits when the market dissolved. With their existing emphasis on analytics, these organizations are far ahead of the industry which can present greater risk early in the transformation, but will eventually give way to an easing of risk retention by lenders. Both the secondary market and sellers can alleviate the costs and time consumption of due diligence processes while still having the ability to ensure the quality of the loans they originated.
Issues such as the conundrum over affordable housing will also be alleviated with credit decisions that will be unique for each borrower. Using AI technology will provide credit decisions based on the attributes presented by the borrower and measure them against the outcomes already included for similar individuals in mortgage performance databases have the potential for allowing more accurate risk assessments and more opportunities for homeownership.
Artificial Intelligence is coming! It is not a question of "if" or "when", but a question of will the industry be ready. Will we not just be focused on the "glitzy" upfront showcase, but on the internal workings of a real AI program. Will we be drivers of the transition from today's work environment to that of the future. The industry needs to spend time and effort on what we believe will occur and begin planning and executing actions to prepare. Without this planned transformation it is likely that this industry will find itself trying to play catch-up or worse. We will awaken too late and find that others who were better prepared have taken over mortgage lending programs in their entirety.
(Views expressed in this article do not necessarily reflect policy of the Mortgage Bankers Association, nor do they connote an MBA endorsement of a specific company, product or service. MBA Insights welcomes your submissions. Inquiries can be sent to Mike Sorohan, editor, at firstname.lastname@example.org; or Michael Tucker, editorial manager, at email@example.com.)