Printer Friendly

Underwriting intelligence.

Underwriting Intelligence

Artificial intelligence may be closer than you think--one company has made great strides in cooperation with IBM to create an expert system for underwriting loans.

The Vision:

Saturday, June 26, 10:00 a.m., in the Metro Realty office. After a tough month, Realtor Phil Garcia finally negotiated a contract on the Cooper homestead. The buyers, Mary and Richard Smith, wait in his office with their pay stubs, W-2s and bank statements. Michelle Petrie of Forever Funding has rushed over from her daughter's soccer match. Petrie plugs her laptop PC into the phone line, accesses the Forever Funding computer, and begins to analyze the Smith's financial resources.

10:15 a.m. Forever Funding's home office computer receives the application data from Petrie's laptop. It orders an appraisal and credit report and determines that the loan might require mortgage guaranty insurance. It checks Fannie Mae and Freddie Mac guidelines and rates and determines the agency of choice. Then the computer transmits the application and credit report data to Mighty Mortgage Insurance's computer, which approves the borrower for a higher loan amount and suggests a single-premium payment plan.

10:18 a.m. Forever Funding's computer transmits the interest rate quote sheet, loan instrument choices, credit history and the mortgage guaranty approval back to Petrie. She reviews the information and advises the anxious buyers.

10:22 a.m. Using the computer-generated authorizations, Petrie qualifies the Smiths for a higher loan amount. She advises them to finance some closing costs including the mortgage guaranty premium. This leaves the Smiths additional funds for renovations. She recommends a seven-year, fixed-rate balloon mortgage, then enters their preference.

10:30 a.m. The Forever Funding computer approves the mortgage subject to the appraisal. Nearby, a laser printer produces the loan application and disclosure documents tailored to the Smith's purchase. They sign immediately. Petrie returns to the match in time to see her daughter score the winning goal.

10:35 a.m. Garcia informs Colonel Cooper that the buyers' financing is nearly complete. Colonel Cooper rebuffs his second cousin, Finley, who had submitted a very low cash offer.

Tuesday, June 29, 2:00 p.m. The appraisal is transmitted to the Forever Funding computer, which obtains final approval from Mighty Mortgage Insurance's computer and confirms pipeline and investor funding. A laser printer in Petrie's office prints the necessary closing documents and mortgage guaranty commitment.

2:15 p.m. The Smiths' and the Colonel's attorneys are amazed to learn that all contingencies are satisfied. The proposed closing date--tomorrow afternoon.

Wednesday, June 30, 4:30 p.m. With all documents in order, the sale is efficiently closed. Petrie has helped another family finance a home and marked another record month. She celebrates quietly--her office has no receptionist, no file cabinets, no stockpiles of blank forms. The silence is broken only by the telephone. Garcia has scheduled two more closings for tomorrow.

Was the preceding account a vision of the future? Or was it closer to being a glimpse at the here and now? For lack of a definitive answer, let us just say it is quite conceivable that a mortgage banker could soon enact many of the scenes from this mini-drama. Much of the groundwork for making it a reality already has been laid, and it remains for mortgage lenders and others to jump on the technology and connect all the links in the transaction chain to make this a more commonplace occurrence.

The Mortgage Data Standards Task Force sponsored by the Mortgage Bankers Association of America (MBA) is presently developing standards for the exchange of electronic data that makes the described scenario possible. Several vendors of credit and appraisal services are developing or using systems to transmit data directly to mortgage lenders.

In this article we will explore another critical technological component of this vision: computers that can make complex decisions. These computers require a technology known as "artificial intelligence" (AI).


In 1987, IBM announced its strategic commitment to artificial intelligence as the technology of the 1990s. IBM established a special expert systems applications department at its Rochester, Minnesota facilities. Its purpose: to assist IBM customers and business partners with AI. The assembled engineers searched for applications to showcase the enormous potential of artificial intelligence.

United Guaranty Residential Insurance Company, based in Greenboro, North Carolina, also identified AI as the foremost technology to deliver faster, more flexible and more comprehensive services to its customers. United Guaranty recognized that significant additional gains in productivity and service required AI technology and began researching AI in 1989.

Several mortgage bankers and mortgage guaranty insurance companies had experimented with limited AI systems. Typically, these evaluated only borrower credit or credit and income. The systems usually issued pre-qualifications or approvals that were subject to a satisfactory appraisal and/or subsequent file review.

United Guaranty envisioned a system that would evaluate all facets of a loan file including borrower income, credit history, the appraisal and the loan instrument. Coupled with advances in electronic communication, this system would ultimately have the authority to approve all but the most difficult loans for mortgage guaranty insurance.

During the initial research period in 1989, United Guaranty was impressed with the expertise IBM had collected. To capitalize on their complementary strengths, the two companies collaborated to develop an AI prototype that could underwrite mortgage guaranty insurance.

United Guaranty and IBM quickly seized the opportunity to combine United Guaranty's expertise in mortgage guaranty underwriting with IBM's expertise in AI. Developing an AI system required a knowledge engineer, that is, a developer of an expert system, who could transform the knowledge in the mind of a human expert into a computer system that could make decisions. IBM initially provided the expertise and then trained a United Guaranty staff person to continue as the project's knowledge engineer. Percy Strader, United Guaranty's expert systems specialist, now has that role, and is the full-time project manager for the AI system.

Currently, United Guaranty has a fully operational expert system that is capable of looking at nearly all of its loans and can render a decision in 10 seconds or less. United Guaranty says the expert system is in its advanced refinement phase at the present time, and thus underwriting professionals are asked to give feedback to the AI program manager to find out how the system is doing its job on individual loan cases. The company characterizes this phase as a "training period" for its new member of the underwriting team, where the expert system is "learning" about the business of underwriting through a step-by-step process of improvements and updated programming.

The expert system evaluates all facets of a loan file including borrower income, credit history, the appraisal and the loan instrument. After a lengthy trial period, it will have authority to approve all but the most difficult loans for mortgage guaranty insurance. In these respects, it is an advancement on previous AI systems that were limited to credit evaluations and borrower prequalifications. These systems often carried a number of contingencies (such as a satisfactory appraisal review) and did not provide unrestricted loan or mortgage guaranty insurance commitments.

The results of early performance testing have been impressive. United Guaranty plans to implement the expert system in early 1992 to assist underwriters with 75 percent to 80 percent of the company's mortgage insurance applications. United Guaranty will be working with mortgage bankers in 1992 and 1993 to establish direct computer-to-computer links in a manner similar to the Forever Funding scenario.

Testing and fine-tuning system performance

Testing has come in two phases, module testing, which took about three months from August through November 1990, and refinement testing, which is currently underway. The results of the initial module testing of a few loan cases produced a working prototype that could evaluate a handful of loan applications.

During initial refinement testing, the prototype of the expert system, running on a personal computer, evaluated 110 loan files. The project team compared each prototype decision to that of the original underwriter, concentrating on those cases where the expert system differed. The knowledge engineers modified the appropriate weights, target values, domain rules or procedural rules to more closely follow the original underwriting decision. These modifications, however, often reversed correct decisions made by the system on earlier cases. For each modification, the team retested all 110 files. The engineers retained the modification if it improved the overall accuracy of the system. If it reduced the accuracy, the modification was sometimes discarded or remodified (even if it correctly fixed the case at hand).

Refinement testing of the prototype against thousands of loan files will produce a production version of the system the company is looking to implement in 1992. Concentrated refinement testing really began last summer--the six months or so before that was spent working out technical details for the actual connection of the hardware.

IBM and United Guaranty believe AI can enhance many decisions in the mortgage origination process. AI can sort through the multiplicity of guidelines, forms and calculations required by secondary market investors, by the FHA, VA and private insurers. Artificial intelligence can fit financing to borrower resources and reduce errors and inconsistencies. It can even promote sales of mortgage-related products to targeted customers. Now we will explore the various phases involved in launching a system using artificial intelligence.

Basic research

At the onset, United Guaranty assembled a team of data processing and underwriting professionals led by Project Manager Strader. The team conducted some basic research to increase its understanding of both the AI technology and the application to mortgage guaranty underwriting. After conducting four months of library research on artificial intelligence, the research team consulted with AI experts and heard presentations by vendors of this capability. At that time, they realized that they couldn't learn all there was to know on their own--they needed the help of a knowledge engineer and turned to IBM for help.

Meanwhile, the project manager visited a remote field office of United Guaranty in Fairfax, Virginia (now located in Falls Church, Virginia) to train as a beginning underwriter for a week. The team found examination of actual loan files to be essential. Specifically, they learned: what an underwriter looks at first and foremost; how an underwriter organizes a file; and how elements relate to each other within the file. By asking questions such as "What if factor X were different?" they soon identified the major risk factors that drive the decisions reached by underwriters. Emphasis of borderline loans often revealed less common, but important, risk factors and decision-making processes.

Setting quality and performance standards

During the refinement testing phase, United Guaranty realized that an AI system with quality and performance standards would help the company's customers generate high quality, conventional, low down payment loans. In turn, this would provide United Guaranty with a reasonable underwriting profit. These standards, then, had to be measurable and practical. They would quantify the business risks of the new system.

United Guaranty's underwriting professionals were sensitive to the fact that the cost of one unexpected foreclosure would offset the profits on as many as 27 good loans. This led to a demanding quality standard: For loans that were clearly an unacceptable risk--the company wanted the expert system to agree with the underwriter professional 95 percent of the time. The remaining 5 percent that the expert system approved in error, therefore, would make up a very small portion of all approved loans, translating into an even smaller increase in default risk.

A high-quality AI system required a moderate initial performance standard. The team expected the first production system to approve the most straightforward 50 percent of applications. These are sometimes called "slam dunks" where there is no complicated credit history and the paperwork is all in order. United Guaranty's professional underwriters would continue to underwrite the remainder. Additional production systems designed to approve more complicated cases are planned for the future.

Underwriting consistency

United Guaranty could not expect the system to be more consistent in its approvals than the professional decision makers it was designed to emulate. To measure the consistency among their underwriters, in order to apply that measure as an acceptable gauge of consistency for the AI system, the team made multiple copies of 110 moderately difficult loan files. Each copy was then independently reunderwritten by an individual different than the original underwriter.

United Guaranty found that there was a slightly higher approval rate for the original underwriting of the loans than for the copied files in the hypothetical situation. In the production environment, underwriters want to help the customer, and in the process, may be more inclined to approve some borderline loans. In the laboratory environment, they may wish to demonstrate stricter adherence to underwriting quality. Both are desirable characteristics of underwriters. The project team and United Guaranty's management agreed to model the AI system on actual field production decisions rather than laboratory examples.

When different underwriters separately reviewed the same loan file, they tended to agree on the loan's acceptability about 85 percent of the time. Dissenting opinions occurred only about 15 percent of the time. Nevertheless, AI provides the opportunity to approve good loans where there underwriting inconsistency may occur. If such loans can be identified and approved, they will produce a moderate source of additional revenue without excessive risk for both United Guaranty and its customers.

Software selection

United Guaranty's specific objectives simplified the selection of software. The team chose a rule-based expert system.

Rule-based systems appear more difficult to program initially than AI technologies that learn by example or history. Yet rule-based systems are easier to create and maintain than traditional programming languages. They naturally provide explanations for decisions. They do not require a historical database. They incorporate and exercise the collective wisdom of multiple underwriting professionals.

There were many rule-based systems available. The team favored products with certain utilities and communications software. They also looked for software that provided hardware flexibility.

A deciding feature that many systems lacked was an interactive development environment. This feature allowed the knowledge engineer to modify rules, observe the effect on individual decisions and trace logic by switching among built-in windows. It reduced the time required to create, refine and test new rules.

IBM had an excellent, highly regarded rule-based expert system called The Integrated Reasoning Shell [TM], but it did not support United Guaranty's operating configurations. With IBM's strong recommendation, the project team chose Inference Corporation's ART-IM software. The expert system development tool played a critical role in module testing. This allowed developers to change rules, rerun and intantly check the new results--all on-line and in "real time."

Hardware selection

Many expert systems, including ART-IM, did not operate in United Guaranty's IBM AS/400 computer environment. That was not a critical drawback because AS/400-based systems support communications with other computers, including personal computers and work stations. IBM engineers tested the prototype and discovered that a decision that required 18 seconds on an IBM PS/2 computer required just 0.8 seconds on the IBM RISC System/6000 work station. This versatile work station supported other leading edge technologies United Guaranty wished to explore.

Expert systems: designing the brain

An expert system is a computer program that makes decisions comparable to those of a human expert. Knowledge engineers model the information-processing skills of human experts such as mortgage underwriters.

Thus far, expert systems have not been easy to build. Communication can be especially difficult due to the technical jargon of both the domain experts (the people whose expertise is to be modeled) and the knowledge engineers. This often creates a "knowledge acquisition bottleneck."

The IBM Applications Technology Development division developed a methodology specifically to eliminate the bottleneck. This methodology divides the expert system design process into four phases:

* outlining the decision structure; * defining the knowledge base

(domain knowledge); * modeling the decision process

(expertise knowledge); * testing and fine-tuning system


Each phase is executed in sequence and requires the solid foundation of the one before.

Two requirements are particularly critical to expert systems. One is having a clear set of objectives. The second is having a good development tool, which is critical in module testing. United Guaranty recognized early on that an expert system would not eliminate costs or revolutionize mortgage guaranty insurance overnight. Instead, the objectives clearly emphasized the delivery of improved service to United Guaranty's customers.

Phase one: outlining the decision structure

During the first phase of designing the methodology by which the expert system would operate, the system developers on the project team undertook a crash course in mortgage guaranty insurance underwriting. They studied underwriting handbooks, procedures manuals, loan package documents and data-entry documentation. They focused on general processes rather than decision specifics.

The most important document portraying the hierarchical nature of decisions is the "conversation sheet." United Guaranty's underwriters prepare this document for each loan application. It distills the typical half-inch-thick loan package into one page of key information.

The conversation sheet revealed the decision process structure for each loan. Two tiers of evaluations led to the final underwriting decision. The first tier evaluated particular loan facts. The assessments of the first-tier evaluations led to four general evaluations comprising the second tier. These evaluations were in regards to:

* ability to make loan payments; * credit management ability; * property marketability (value); * loan instrument risk.

Therefore, the results of these four second-tier evaluations determined the final decision on whether or not to approve the loan.

Knowledge engineers often examine the process in reverse--from the final decision backward to the raw facts in the loan file. They might initially determine a second-tier general evaluation such as the first one: the ability to make loan payments. Next they might identify the first-tier detailed evaluations that led to it:

* the ratio of mortgage payment to

income (front ratio); * the ratio of total debt payment to

income (back ratio); * employment stability of the

borrower and co-borrower. * types of income (wages, overtime,

bonuses, commissions) and the

security of each.

The conversation sheet also revealed that most evaluations measured relative strengths and avoided binary (yes/no) decisions. Values exceeding underwriting guidelines did not produce automatic rejections. The system assessments, therefore, would not be "good/bad" but would be a numerical rating on a scale. Knowledge engineers often use a scale of -1.0 to +1.0 instead of the familiar scale of 1 to 10, but the idea is the same.

Phase two: defining the knowledge base (domain knowledge)

The knowledge engineers had to learn how to process or map the raw data available in loan files (income, mortgage payment, sales price) into key ratios and statistics. These would become the actual values and target values in the evaluations in the structural model designed from phase one.

Domain (mapping) knowledge defines how to process raw information rather than how to make decisions. The source for this knowledge was United Guaranty underwriting guidelines and standards of practice.

Phase three: modeling the decision process (expertise rules)

Acquiring expertise from underwriters required lengthy collaboration. The engineers asked the experts to underwrite loan files out loud. They frequently asked for details about the underlying reasoning. Close cooperation and trust were essential. At some point, every underwriter in the company assisted with the project.

This third phase entailed capturing expertise by setting target values, developing standardization formulas and assigning weights. These numeric and algebraic components became the means by which the expert system really "learned" the underwriting expertise.

The process of assigning weights began with underwriters ranking firstier evaluations in order of importance. These rankings became the basis for assigning numeric weights. For example, underwriters ranked the back ratio higher than the front ratio in determining the borrower's ability to make the loan payments. Therefore, the back ratio received a higher weight in the evaluation rule.

Numeric weights allowed the various first-tier assessments to be combined into second-tier general assessments, and the latter into a final decision. The engineers would refine the weights many times during the testing phases. To facilitate frequent adjustments, the weights would be placed in data tables as had been done for target values.

Case studies led the project team to another important discovery: expert underwriters develop shortcuts to reduce deliberation time. One shortcut was to look for excessive values of certain "hard" or inviolable guidelines. The expert system would incorporate this strategy with high-priority or "hard" rules. The triggering of a high-priority rule would automatically refer the insurance application to a professional underwriter. An example of a hard rule for standard underwriting guidelines is that the property be owner-occupied.

Some risk factors rarely influenced decisions unless several occurred together in a loan file. The knowledge engineers created "soft" rules, each triggered by the presence of a different risk factor. Examples of soft rules are a low level of cash reserves after closing and an occasional late payment on installment debt. Unless there were other weak features of the loan, the system would have the flexibility to approve such loans. When three or more of these "soft" rules were activated, the case would be sent to an underwriter professional. The triggering of two or fewer "soft" rules would produce no effect.

Obstacles in capturing data

Refinement testing encountered a common obstacle--United Guaranty's data-capture system, designed in the early 1980s, had not anticipated the requirements of expert systems. Some important loan facts had not been entered into the existing system, such as the value of the comparables, the square footage of the home and past delinquencies and foreclosures.

Once the project team identified the missing data elements, a special team of system developers began redesigning the data-entry system. This was a major project around February 1991 and the new design of the data-entry system was completed in October. The redesign was important to avoid a common reason for expert system failure: forcing an expert system to run on inadequate data. This would have been comparable to asking an underwriter to make a decision without the full loan application package.

To design the new data-entry system, the project team had to create new data. Here the ability of the expert system to run on different computers proved crucial. Using a personal computer, the project team bypassed the production system to encode all necessary data from 110 loan files. The team specifically selected cases that covered a broad range of risk characteristics. Massive refinement testing of nearly a thousand cases began once the new data-entry design became available in October.

The vision is now

IBM and United Guaranty will demonstrate the jointly developed prototype to interested parties. Depending on the desired response time, the mortgage underwriting system can reside either on an IBM RISC System/6000 work station or an IBM PS/2 Model 70. Each case is down-loaded from an IBM AS/400 host computer via TCP/IP (communication protocol software) and evaluated. The underwriting decision is transmitted back to the host. The communication protocol allows for host computers other than the AS/400.

With United Guaranty's plans to fully implement its communications with lenders in 1992, the active use of artificial intelligence in mortgage lending may not be as visionary and futuristic as you might have thought. It is one of many advanced technologies that is currently reshaping the way this business will be done.

While the system only underwrites mortgage guaranty insurance, IBM and United Guaranty believe there may be wider applicability in the mortgage lending community. IBM and United Guaranty are promoting this technology as a way to improve service and explore opportunities with innovative customers.

Daniel Walker is vice president of actuarial and economic research for Greensboro, North Carolina-based United Guaranty Corporation, and directs the company's AI development team. Marc Gluch-Rucys is a senior associate programmer in the expert systems applications department at IBM, Rochester, Minnesota.
COPYRIGHT 1991 Mortgage Bankers Association of America
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 1991 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Title Annotation:includes related article; United Guaranty Residential Insurance Co. and International Business Machines Corp. develop an artificial intelligence system
Author:Gluch-Rucys, Marc; Walker, Daniel
Publication:Mortgage Banking
Date:Dec 1, 1991
Previous Article:The new origination game.
Next Article:Developing Incentive Compensation Programs for Mortgage Lenders.

Related Articles
The brave new world of artificial intelligence.
Artificial intelligence on a budget.
Artificial Intelligence Frontiers in Statistics: AI and Statistics III.
Al technology inspires NewsRx to move beyond health care, spinning off new media company.
United Guaranty's new look. (Profile).
Insuring dreams: Mortgage guaranty insurers are successfully navigating a mixed environment of low interest rates, increasing unemployment and...
Best's rating changes.
Mortgage insurers' vintage hangover: mortgage insurers are going to have to boost loss reserves as carriers deny claims based on fraudulent behavior,...

Terms of use | Copyright © 2017 Farlex, Inc. | Feedback | For webmasters