Printer Friendly

The brave new world of artificial intelligence.

Welcome to the strange new world of "mentoring," where underwriters take computers under their wing and teach them everything they know. That's the future in mortgage lending, and it's like trying to download an underwriter's brain and "compassion."

No longer in mortgage lending is it standard for all loan officers to make calls with nothing more than rate sheets, blank applications and a pen. Today, a growing number of lenders are arming their retail sales force with laptop computers. And some are planning to use artificial intelligence (AI) software to underwrite mortgages on those laptops, while you wait--thus providing instant loan approvals at the point-of-sale.

The list of major mortgage lenders aggressively moving ahead with artificial intelligence capabilities in their origination operations is impressive--and growing. Household International, Prospect Heights, Illinois, is using an "expert system" to underwrite more than half its retail and wholesale mortgage loans. Norwest Mortgage, Inc., Des Moines, ARCS Mortgage, Inc., Calabasas, California, and Countrywide Funding Corporation, Pasadena, also are currently developing similar systems. Home Savings of America, FSB, Irwindale, California hopes to have a "sophisticated credit scoring" system in place early next year, according to a top executive. A source within Fleet Mortgage Corporation, Milwaukee, says that institution is interested in artificial intelligence, and industry sources say that Fannie Mae also has licensed AI software.

Growth of automation

It's no secret that technology is rapidly and irrevocably changing mortgage lending. Consumers enjoy faster processing, and lenders find they can handle variations in volume more easily now than ever before. In addition, lenders enjoyed better profits on this year's business, because they did not have to hire as many new production staff to handle the increased business, as they did during prior spikes in volume. Streamlined processing requirements adopted by Freddie Mac and Fannie Mae, plus automation integrated into lending operations since 1987, combined to help the work flow during the current refi boom.

Computers now can hold the necessary information regarding which documents need to be drawn up for closing different types of loans. With this kind of assistance, a temporary worker, lacking a mortgage background, can be hired to help handle the additional volume in a closing department. With proper supervision and a good automated system, a clerical worker can make loan closing more productive immediately.

Servicing takes the lead

Further evidence of the heavy inroads made by technology in all aspects of mortgage banking is readily apparent in the servicing area. Instead of preparing tax and insurance bills and sending them out, many servicers exchange computer tapes. Tape-to-tape insurance renewal started at Mortgage Guaranty Insurance Corporation (MGIC) in 1978, says Richard Lembach, manager of automated services. Computers can quickly sift through thousands of borrower accounts to see who needs to pay, disburse the funds and then update records on the computer tapes.

Because automation lends itself to repetitive tasks, servicers have found many uses for technology. Today some servicers are starting to scan loan documents into imaging systems, so that customer service representatives can immediately find the information they need to answer an inquiry. When used on a computer network, people in different departments can look at the same document simultaneously.

Faster service and reduced reliance on staffing should translate into better profits and happier customers. Yet not all technology works, and there is a cost to being on the cutting edge. For instance, computerized loan origination (CLO) networks have been tried numerous times by large companies, with mixed results, at best.

Loan origination aids

But loan officers for some firms today, with the aid of technology, can grant approval at the time of application, conditioned only upon an appraisal. Doing so provides an efficiency the mortgage market has never quite experienced before.

Today, sales representatives for B.F. Saul Mortgage, a subsidiary of Chevy Chase Federal Savings Bank, have "well over half of their applications approved" in a few seconds, says Chief Information Manager Robert Spicer. Maryland-based B.F. Saul's MortgageVision system allows originators to electronically capture credit report information by having laptop computers dial into credit data repositories immediately after taking a loan application.

Credit and application information then is sent to an artificial intelligence system, which either approves the file or refers it to a human underwriter. If the loan is approved by the AI system, the loan officer can let the happy borrowers know immediately. B.F. Saul set up Mortgage Vision three years ago, according to Spicer.

Spicer also notes that the system allows "higher volume, a lower cost of originations and a raised level of service." He adds that capturing information in a Realtor's office or a borrower's home allows the loan to "move to servicing electronically," without requiring any rekeying of borrower data.

Changing the culture

Right now, however, many loan officers are still adjusting to computers, voice mail and car phones. Furthermore, some large mortgage companies were built up around the idea that loan officers fill out 1003s and then give them to clerical workers to type up or enter into the computer.

When BancBoston Mortgage Corp., Jacksonville, Florida, decided in 1988 to put its loan officers on laptop computers and require them to meet minimum production standards, 80 percent of the firm's originators left within a year, according to Senior Vice President Thomas Palmer. Yet, BancBoston achieved efficiencies by tying the laptops into its automated processing system and by encouraging its originators to stay out in the field, rather than in the office.

Additionally, the system was introduced into the offices of large Realtor clients. Branch production costs dropped during the next year for BancBoston Mortgage, from 199 basis points to 103 basis points, according to Palmer.

Even so, many loan sales reps continue to work without much benefit from automated systems. According to one industry estimate, less than two-thirds of all mortgage brokerage firms use technology to speed loan processing or credit report ordering. Can they be coaxed into using automation more?

Freddie Mac is offering all lenders its "Gold Connection" software, which allows them to get prices and commitments without talking to another person. Lenders will be able to dial directly into Freddie's computers from their PCs and get a loan locked without using the normal commitment phone line.

Other computer uses

For years, computers have been helping underwriters make decisions. Miller Warren, Jr., senior vice president of information systems at GE Capital Mortgage Corp. in Raleigh, North Carolina, notes that the firm has been using "computer-assisted underwriting" for two decades. GE's system warns human underwriters about ratios, urges them to consider the whole package, and points out overriding factors, which would prompt a rejection, such as loan-to-value (LTV) ratios over 95 percent.

In the 1980s, PMI Mortgage Insurance Company, San Francisco, developed a statistical-based model that predicted problem loans by comparing "a borrower profile and some property elements" with the company's database, says Vice President of Customer Technology Pat Mikel.

Although looking just at selected loan file elements is adequate when underwriting some mortgages, others require a synthesis of information about the borrower, loan and appraisal. For instance, B.F. Saul's system "mines the credit report data," says Spicer, drawing more than 60 percent of the underwriting information from this source.

Even if an application isn't approved by the AI system, it shows areas of concern, allowing the underwriter to zero-in on problems. Just by clicking on a computer "mouse," an underwriter might quickly see that if the applicant pays off one debt, then he or she will be within guidelines. Artificial intelligence "is really paying off," according to Spicer.

System choices

AI software "is like an empty brain," says one industry expert. Lenders attempt to get it to mimic the thinking process of their best underwriter.

"Judgment or case-based" systems learn underwriting by being provided application information on many loan samples. Then they are told whether or not the loan actually was approved. After seeing enough examples, the AI system learns what causes a loan to be approved or rejected. For instance, B.F. Saul used "tens of thousands of cases to develop the model," Spicer says.

George Lowery, vice president and chief information officer at ARCS Mortgage currently is feeding the firm's AI system with past loan files. When the system doesn't understand a loan, it asks for help from its "mentor"--a role generally played by the lender's head underwriter. Mentors work with systems experts to teach AI software.

ARCS will spend a year developing its AI system before an anticipated February 1993 debut in branch offices. Then it will spend another year being tested alongside human underwriters before being allowed to make decisions on its own. Eventually, Lowery hopes to be able to increase loan volume without having to add underwriting staff, and to avoid the inconsistency resulting from the "good days and bad days" experienced by all human workers.

To make a decision, ARCS' artificial intelligence system looks at a borrower's credit and employment history, as well as the home's appraisal. Data will be transmitted each night to the firm's mainframe computer, and underwritten in two-and-a-half seconds, Lowery says. Acceptances will be faxed to the branch the next morning, while "anything questionable will be forwarded to the teaching model" to further refine the system.

Lowery says that the AI system will "look at the factors in the loan package to come to a judgment and a degree of confidence rating" for that decision. But he notes that "if you teach it incorrectly, it will give wrong answers." In addition, nothing works if loan application information is not correctly keyed into an expert system. For that reason, some firms build hundreds of data checks into the software, to help spot incorrect entries.

As common as spreadsheets

ARCS and several other lenders are developing AI systems on software licensed from Cybertek-Cogensys Judgment Systems in Dallas. President Joseph Filoseta predicts that AI systems one day will be as prevalent as spreadsheets. In fact, an AI system can be run on a PC. Furthermore, firms using AI can "capture intellectual capital," Filoseta contends. "No one retires, dies or goes to a competitor."

Filoseta claims that a lender originating at least $750 million annually will find the system pays for itself in a year. Industry sources say the cost of licensing AI software is in the hundreds of thousands of dollars.

Once an expert system is developed, Filoseta suggests continuing the mentoring process one to two hours a week to monitor market changes. B.F. Saul's Spicer has found that MortgageVision requires one part-time worker to maintain the system.

How closely do these systems mimic their mentors? Greensboro, North Carolina-based United Guaranty Corporation's "rule-based" system, developed with IBM, currently is functioning invisibly," as a test to see how its decisions compare with human underwriting, says Daniel Walker, vice president of actuarial and economic research. About 20 percent of the time, there is a disagreement, and United Guaranty is studying those cases to determine reasons for the discrepancy and their "impact to the corporation," Walker adds.

However, AI isn't cost-effective unless it is used instead of human underwriters much of the time. Countrywide Funding Corporation Executive Vice President Ralph Mozilo says 40 to 50 percent of loan applications must be approved by the system to justify its costs. Mozilo is hoping for an eventual level of 70 percent approval, which is in line with targets held by other lenders.

Conservative approach

Norwest Mortgage Inc. is working to mentor an AI "system that will do all types of loans on a pre-qualifying basis," says Phyllis Bement, vice president of national underwriting. She notes that an AI system can't replace FHA delegated underwriters.

"It is a very long process," Bement adds. "Eventually we will let the system decide approvals on some loan products at some LTVs." She explains that Norwest will not use AI to decide on loans of 90 percent LTV or higher, or on adjustable-rate mortgages. Bement anticipates the system initially will underwrite conventional loans with LTVs of 70 percent or less.

"There are so many variables," she adds. "It's impossible to teach a system all of them." Bement explains that her underwriters work on 14,000 loans each month. "Even if we entered them all in the system, you can't teach it every possible difference." Norwest currently is teaching its expert system to analyze credit, and later will add an appraisal component. Bement wants the system to analyze credit and appraisal data separately and then as a whole picture.

Norwest has been working on its AI system for eight months, and "will marry it to the front-end system in the first quarter of 1993," adds Bement. She adds that it will be used only to pre-qualify buyers during "whatever period is necessary to be sure it isn't making decisions that will jeopardize Norwest stockholders. We aren't under any type of pressure to beat the competition."

Bement notes that simply handling pre-qualifications for originators will be a big help. "We're doing $100 million a day. Even if they call underwriters just on 10 percent of those, it would save a lot of phone time" to pre-qualify on a computer, she explains. Bement speculates that in the future Norwest's AI system could be linked with the computer systems of big clients, and also be used by sister companies, such as Norwest Financial. She adds that Norwest's expert system is just "a piece of a bigger project."

But for now, Bement hopes AI will keep her from having to hire another 80 underwriters during the next 18 months. In addition, the time freed up for her 150 current underwriters will allow them to do more training of loan officers and processors, plus meet with Realtors and builders.

Changed processes

Nothing as revolutionary as artificial intelligence can come into a company without encouraging shifts in operations. Senior management at FBS Mortgage Corp., St. Paul, is thinking of ways its new judgment-based system can add to loan originations and processing. In order to help underwriting staff manage during peak volume--whether it's this year's refi boom or the last two weeks of any month--FBS Mortgage will use the system it's currently mentoring "to review loans and point out |areas of~ concern, so underwriters don't have to go through every piece of paper in the file," says Underwriting Analyst Debra Ringler.

Following such an approach, also "moves underwriting questions |up~ earlier" in the process, Ringler notes. "At the time of application, we might decide not to go forward." Being able to grant approvals or show what needs to be verified at the time of application is viewed as a major benefit by FBS, says Senior Operations Analyst Tracy Howe.

Currently, FBS Mortgage has no roll-out timeframe for its artificial intelligence system. Loans are being input now, and the system later will be given identical loan applications to see if it agrees with the original underwriting decision, adds Howe.

During the mentoring process, Ringler's time is dedicated to the AI system. She notes that working exclusively with one mentor who decides how to break underwriting decisions "into bite-sized problems" is important. Ringler adds that using a committee of underwriters tends to confuse the computer.

Some AI users later "wish they had designed their system differently," Business Analyst Bob Stokke, also with FBS Mortgage says. Ringler explains that "the art of artificial intelligence lies in knowing when to use tables to give explicit guidelines, and when to leave the situation less-defined." Using tables allows the system to be changed easily, but Ringler says too much reliance on them means a lender can "lose the artificial intelligence benefit."

She adds, "it's very unlikely two lenders' systems will be the same," even though they are both originating identical, conforming loans. Allowing room for the system to incorporate new loan products within its learning framework also is important, Ringler says.

Rule-based systems

Judgment-based artificial intelligence systems--also known as intuitive systems because they infer what is important--make loan decisions without asking for, or generating, any explicit rules. Instead, a new application is related to loans the system has been shown were approved or rejected in the past. Patterns are looked for, rather than simple answers.

Yet, the system is able to see some decision-making patterns that have escaped its mentors. For instance, underwriters have been surprised when an AI system has told them that some of the factors it was initially told were part of the approval process, weren't being used to make decisions.

Lenders using judgment systems say the software keeps learning because it always asks for help when it comes to an example that it hasn't seen before.

However, some lenders prefer to have more control over what their system is learning. Rule-based artificial intelligence uses a series of statements telling the system, "If you see this, then do that," says Mike Meade, senior vice president of information services at MGIC. He adds that it takes longer to define decision rules through that type of programming. But Meade says that changing underwriting requirements is easier with a rule-based system, because a few new rules can be inserted, rather than having to re-teach a judgment system through new examples.

Rule-based systems don't proceed step by step like a decision tree does, notes United Guaranty's Walker. Instead, a system will use hundreds of rules about LTV, debt ratios and other factors to formulate a decision.

One way to combine traits of both systems is by giving a judgment-based system a series of tables defining basic loan parameters. Certain ranges of values for LTV, for instance, would be defined either as excellent, marginal or poor. If a lender then wanted to change some basic underwriting assumptions, all it would need to do is alter these tables.

Putting it together

Countrywide Funding's Ralph Mozilo expects to roll out an AI system by December 1992 for use in the firm's retail and wholesale origination operations. Initially, the parameters of Countrywide's rule-based system will be drawn narrowly, so that only 35 percent to 40 percent of all applications will be approved without going to an underwriter.

Mozilo says that close to 200 pieces of data will be taken from the loan application, appraisal and credit reports. Yet, the system can make underwriting decisions in 10 seconds to 15 seconds when loaded on a 386-PC with 8 megabytes of memory.

Artificial intelligence has marketing benefits, as well as the more obvious underwriting uses. Being able to tell Realtors that loan approval often can be instantaneous will be helpful, as will providing mortgage brokers and correspondent lenders with delegated authority to approve loans based on AI decisions. Consumers could even fill out an application on their home PC, get it approved and lock in a rate on-line, Mozilo says.

He also envisions artificial intelligence helping spot irregularities that might indicate fraud "that a processor or underwriter might not see." Underwriting decisions also can be tracked to see how the loans perform.

Fannie Mae, Freddie Mac and mortgage insurance companies are taking an interest in the Countrywide system, according to Mozilo. Countrywide is discussing electronic loan transmission and faster payment with the agencies, as well as seeing if they would certify that mortgages approved by AI "would not be eligible for repurchase due to the underwriting decision," notes Mozilo.

One added benefit, according to Mozilo, is that the system is "color blind. There are no rules regarding race or sex." Currently about 600 to 700 rules are on Countrywide's AI system. But to be able to accommodate a wide-enough variety of borrower situations and loan types to approve 70 percent of all applications, about 2,000 rules will be needed eventually, says Houman Talebzadeh, director of artificial intelligence at Countrywide.

An average of 30 percent to 40 percent of all rules will be used to make a decision on any one application, he explains. Whenever the system doesn't approve a file, it is sent to an underwriter with suggestions for resolving the problems.

As with other large lenders, Countrywide's use of AI is just part of a mortgage automation package that starts with a front-end system designed to eliminate any errors from the application. Additionally, Countrywide soon will allow correspondents and mortgage brokers to use the AI system. Brokers also will have on-screen access to status reports on loans being processed, in order to reduce phone messages.

For the future

However, it isn't necessary to develop proprietary systems to make use of artificial intelligence. For instance, Firstar Home Mortgage Corp., Milwaukee, is licensing an AI system for use with its off-the-shelf mortgage origination software. AI software firms also are starting to develop modules for use by smaller lenders. Some are derived from fully mentored systems already built by larger mortgage firms. Combinations of systems also are being used. For instance, MGIC now is working with judgment-, rule- and statistical-based AI systems. A statistical-based approach can be "a means of validating a judgment-based model," says Nick van der Schalie, director of mortgage products at HNC, Inc. in San Diego. Providing statistical assessments of risk allows lenders to better understand the decisions that judgment-based systems make.

HNC's "neural network" system applies varying weights to elements in a loan file in order to predict performance. One application of this AI system might be to use it to assess which mortgages will become delinquent and deciding how to approach them in advance.

For a growing number of lenders, artificial intelligence soon will be playing a role in making the route from application to servicing portfolio faster, less expensive and closer to being seamless. The AI trend probably has more potential to change mortgage lending than companies developing it currently will admit. "AI doesn't cost a lot to implement," notes Norwest's Bement, "but it does have a cost to be three years late."

Howard Schneider is a freelance financial writer based in Ojai, California.
COPYRIGHT 1992 Mortgage Bankers Association of America
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 1992 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Title Annotation:mentoring in mortgage lending
Author:Schneider, Howard
Publication:Mortgage Banking
Article Type:Cover Story
Date:Oct 1, 1992
Previous Article:The new thinking on prepayments.
Next Article:In the year 2002.

Related Articles
What's hot in technology.
Artificial intelligence on a budget.
The big picture.
Reengineering quality control.
Pioneering new channels.
Park Avenue Commercial Group.
Big appetite. (Portfolio).
Everyone's online.

Terms of use | Copyright © 2016 Farlex, Inc. | Feedback | For webmasters