Printer Friendly

Here and now high-tech.

many mortgage processes. Yet frequently, some of the most high-tech origination systems we read about still exist as someone's vision of what will be, rather than as routine workhorses functioning at full speed.

At San Francisco-based PMI Mortgage Insurance Company, the technology strategy discards the Star Wars-type approach in favor of practical, real-world solutions. This strategy is evident in our development and implementation of an integrated automated underwriting system that exploits the capabilities of artificial intelligence (AI) technology and predictive statistical models. By taking advantage of the latest AI technology to expand and improve its successfully deployed statistical model, PMI has been able to further enhance customer service, increase productivity among underwriters and improve the quality and consistency of risk-based decisions.

These benefits, however, were not gained by merely waving a magic wand. Years of analysis, testing and more testing - as well as a little sweat and tears - went into the development of the integrated system to ensure that the technology delivered a pragmatic business solution. That goal became a reality in March of this year as all of PMI's 20 offices nationwide went into full production with the automated underwriting system.

In the beginning...

In 1986, PMI, in conjunction with the Allstate Research Center in Menlo Park, California - the "think tank" of the Sears financial companies - began development of the Automated Underwriting Risk Analysis system(SM) (AURA(SM). The system employs statistical predictive models to predict the likelihood of default by a mortgage borrower.

AURA consists of two interactive statistical predictive models: claim and risk. The models interact to assign each loan a risk score from 1 to 100, with the higher number representing the higher risk.

Fully implemented in 1987, AURA contains more than 650,000 borrower profiles and provides a predictive score for all of the mortgage insurance applications PMI receives. Today, of PMI's total approved business, 60 percent of applications have been approved with AURA. By freeing underwriters to deal with more complex credit packages, as well as allowing them to spend more time familiarizing themselves with their markets, AURA has greatly improved PMI's underwriting efficiency.

Building an electronic foundation

Shortly after the implementation of AURA, PMI began developing an electronic Lender Network, the objective of which is to provide customers with a paperless, optimally efficient mortgage insurance transaction. PMI worked with B.F. Saul Mortgage, a subsidiary of Chevy Chase Bank, Bethesda, Maryland, to create a prototype of the network, which allows mortgage loan officers to electronically submit applications for mortgage insurance to PMI. Loan officers receive an approval decision on their laptop personal computers within seconds, while referrals - applications unacceptable to AURA - are reviewed by a PMI underwriter within minutes. This prototype Lender Network has been successfully used since June 1990.

In order to fully implement an entirely paperless transaction, however, we had to provide the next automation link after AURA - the appraisal analysis. The AURA analysis is limited to the borrower and the borrower's credit history, so that even when AURA approved a loan, a human underwriter still had to evaluate the property being used as collateral for the loan.

In 1989, PMI began exploring ways to automate the real estate appraisal process and to convert it from paper to data elements, which would allow us to achieve our fully electronic transaction. An expert system was determined to be the best solution.

After investigating and evaluating several software providers, we decided to license judgment processing software from CYBERTEK-COGENSYS Corporation of Dallas. The judgment processing software is an expert system that is based on inductive technology. The software is based on the premise that pragmatic experts make decisions on the basis of intuition, experience and knowledge.

In the inductive environment of judgment processing software, the computer assumes the role of apprentice. As it observes the performance of its "mentor" - the person whose expertise the organization wants to capture and leverage - the system creates a dynamic, judgment model. In other words, the software is able to emulate the decision-making capabilities of a company's best and most experienced people without requesting or generating any explicit logic rules.

Judgment processing software differs from the traditional approach to expert systems in that no knowledge engineering process is required. The knowledge engineering process requires that a technician, usually called a knowledge engineer, interviews and observes the expert in his or her area of expertise. Knowledge engineering can be difficult and time consuming for the many experts who may be more intuitive than analytical in their decision-making practices.

Once we opted for this software, PMI's next step was to assign a mentor to help teach the appraisal system. The mentor tackled the none-too-simple task of determining which of the myriad data variables - or decision factors - associated with appraisal decision making were required to arrive at a good decision about the appraisal.

How the system learns

Under judgment processing, the underwriting decision factors are phrased as a series of questions. In defining the questions, the mentor indicates how individual responses are to be categorized. For example, in mortgage insurance underwriting, certain ratios, such as debt-to-income or loan-to-value (LTV), are calculated to evaluate the application. Whether the LTV is 81 percent or 83 percent is not as relevant to the underwriter as the fact that both are more than 80 percent but less than 85 percent. The "between 80 percent and 85 percent" category is called an intermediate judgment.

After defining questions, valid responses and relevant categories, the mentor enters a series of examples and makes a decision or interpretation for each one. The software observes the interpretation and the intermediate judgments associated with each interpretation and stores this information as a situation in a consistent, logical model called the judgment base. As the base develops through the addition of more situations, the system acquires the ability to: * identify relevant past experience; * compare situations and measure

their degree of similarity; * predict how the mentor would interpret

a particular example; * report, on a scale of 1 to 100, a level

of confidence that the mentor would

agree with the decision generated.

During this teaching process, the mentor either accepts the system's interpretations or changes them. In either case, this provides a learning experience that enriches the extrapolation abilities of the judgment Processor, the component of the software that emulates the decision-making logic of human experts.

After determining which data variables, or decision factors, are required to build a model for decision making, then problems that incorporate all of the decision factors are defined. Forming problems enables complex decision processes to be separated into manageable blocks.

After two months, PMI had collected 140 data variables to form a very good appraisal model. However, we learned from feedback generated by the judgment processing software that many of the data variables were irrelevant to the decision-making process. During the nine-month development and analysis phase, the software was able to tell us what data variables are logically important in making valid appraisal decisions. Further teaching and testing during the next seven months dramatically improved the appraisal model. The number of data variables thought to be required to underwrite the appraisal was whittled down from 140 to 36. In addition to providing better appraisal decisions, the reduction in the number of data variables also enhanced greatly the model's ease of use. Users now have to enter only 36 variables, rather than the 140 that were originally thought to be important and which would have required entering screen after screen of information.

Expanding our vision

About six months after PMI started the development of the appraisal model, we realized that an additional judgment model would be required. Because appraisals deemed unacceptable by the appraisal model were referred by AURA - regardless of the AURA borrower credit profile - to a human underwriter, we determined that we were losing some internal operating efficiencies.

We noted that many appraisals determined to be unacceptable by the system were often approved by the human underwriters after they concluded that the borrower profile and the credit history outweighed the marginal aspects of the appraisal.

The second judgmental model, the referral model, combines the borrower profile, the credit history and the appraisal data into one evaluation that determines if positive attributes offset any negative aspects. The result is that we furthered our goal of serving PMI's clients as fully as possible, as well as streamlining the underwriting process.

We learned so much from the development of the appraisal model that the referral model took only six months to develop. In addition, the process was quickened because much of the data used to train the model had already been captured in electronic form during development of the appraisal model.

The crucial test

PMI then took the next step - validating the appraisal model. PMI's management recognizes how crucial it is to test any system thoroughly - and in several environments. Such testing ensures that the objective - to automate risk decisions and increase efficiency without jeopardizing quality - was achievable. After considerable testing in PMI's San Francisco home-office environment, in October 1990, we set out to prove the technology in a production environment. Our San Ramon, California field office was selected as the first test site.

The appraisal model was tested for six weeks, during which time more than 500 real estate appraisals were evaluated. During the validation phase, human underwriters concurrently underwrote the appraisals in order to determine the validity of the automated model's conclusions. Results of the San Ramon validation were excellent: the system approved 75 percent of the appraisals submitted, and of these, the field underwriter disagreed with only one.

In February 1991, PMI set up a second beta test in Tampa, Florida, a market which differs significantly from San Ramon. During the four-week test, the model evaluated 500 appraisals and approved 68 percent of them. As in San Ramon, only one of the approvals was unacceptable to the field underwriten

In June 1991, both the referral model and the appraisal model were tested in San Ramon and Tampa. The integrated system came through with flying colors. The referral model increased the system's efficiencies as hoped, and compared with the previous beta tests, the number of approved appraisals increased in San Ramon to 87 percent and in Tampa to 72 percent.

PMI began implementing the referral and appraisal models throughout the field offices in September 1991. The roll-out was conducted slowly and carefully, in keeping with our policy of preventing any potential disruption in service to PMI customers. By March 1992, all 20 field offices were in live production. Figure 1 demonstrates the flow of data throughout these processes.

PMI's underwriting staff have reported that the models are easy to use. All that is required is the entry of the 36 data variables and the pressing of a command key. A decision is processed in two to seven seconds. These processing times will improve when minor hardware changes are made later this year.

It is important to note that PMI management made sure to prepare the users for the introduction of the new technology. We learned during implementation of the AURA system that a change in how people work can pose a threat to them, so PMI held sessions and distributed newsletters that informed the users about how the models would deliver benefits, not jeopardize their jobs.

Real-world results

Today, the judgmental models are approving approximately 75 percent of real estate appraisals without human intervention. As PMI performs modifications to accommodate unique markets, this percentage is expected to improve even further.

If the judgmental models encounter a unique appraisal case that they have not been exposed to before, they provide an educated response, then provide the assigned PMI mentor a teaching task. The mentor then interfaces with the models on a daily basis to either teach them how to respond to the new situation or to confirm that the original response given by the system was indeed correct.

This process provides information to the judgment base on a continuing basis, thus ensuring that the models are constantly learning and remain dynamic with market changes. By May 1990, the system was taught 23,000 unique situations of appraisals and referrals and encountered an average 1,000 situations per day.

With the automated underwriting system, PMI's senior underwriters are available to spend more time on highly complex risk decisions. In addition, all underwriting staff have more time to develop client relationships, provide client education and sharpen their knowledge of their local markets. Peak business periods are handled efficiently, and the system allows PMI to achieve its primary goal of giving superior service to its customers because it helps to ensure fast responses and consistent decisions.

Other benefits with the system include maintaining a level workforce with less of a need to hire and fire employees as a result of seasonal market conditions. Further, the system enhances the productivity of employees and allows underwriters to concentrate on borrower, credit and appraisal scenarios that warrant closer scrutiny. And, from a portfolio management perspective, the system's quantitative scoring of portfolios allows for improved analyses and reporting.

More promise for the future

PMI continues to search for new "problems" inside the company that can be solved with the Judgment Processor software. In addition, we are positioned to meet our customers' needs through the expansion of the electronic Lender Network to achieve the single-entry, paperless transaction.

No matter what technological frontier PMI explores next, the same painstaking approach taken in the development of the automated underwriting system will be used.

THE

NUTS

AND

BOLTS

The appraisal and referral models are deployed in a client-server, IBM Toker Ring local-area network (LAN) environment, using a personal computer as the decision server Mature judgment bases are accessed through a program called the Decision Delivery Module (DDM), which acts as the decision server and responds to requests for decisions. Users access the models via field office terminals.

The DDM decision server monitors a shared disk directory on the network for transaction records being written into a disk file called the decision request queue. Each decision request record must contain the data elements that the mentor has identified as required for the decision model being accessed. The DDM returns a decision to a response queue.

The DDM also monitors the confidence levels of the responses. If a consultation with the mentor is required, the transaction can be routed first to an underwriter, so that the transaction can be completed and an additional learning experience be provided to the judgment base on a teaching PC.

A direct data communications link from the PC LAN running the judgment processing software to PMI's IBM AS/400 computer was implemented using IBM's LU6.2 cooperative processing protocols. The AS/400 houses the AURA analysis.

Pat Mikel is vice president for customer technology and Terri L. Baker is director of product support in customer technology at PMI Mortgage Insurance Company, San Francisco.
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:includes related article; PMI Mortgage Insurance Co. adopts advanced technology for mortgage processing efficiency
Author:Mikel, Pat; Baker, Terri L
Publication:Mortgage Banking
Date:Jun 1, 1992
Words:2476
Previous Article:In search of real mortgage bankers.
Next Article:An operational redesign.
Topics:


Related Articles
The MI technology race.
Private mortgage insurance.
Ellie MAE and PMI form alliance. (Tech Newz).
Mortgage guaranty.
Making the most of a digital world: here are some tips for how lenders can make the most of a digital business environment, improve customer service...

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