A neural network - could it work for you?

A neural network--could it work for you?

The cream of the artificial-intelligence crop today is the neural network--systems that learn from their experiences. Read how you can use the sophisticated technology to improve your financial decision-making. The power of the computer is about to take another quantum leap, but this time the change won't be measured in more bytes for the buck or processing speed. It will be something entirely new: highly sophisticated computer systems with the ability to learn and to utilize accumulated experience to make specific, sound business decisions that rival, and perhaps surpass, those of a living, breathing human being.

Called neural networks, such artificial intelligence systems are built around concepts similar to the way the human brain's web of millions of neural connections (synapses) are believed to work together to identify patterns, learn, and reach conclusions.

Considered by some to be one of the most important technological advances of the last 10 years, neural network systems already are at work in the world of finance and elsewhere. They are particularly applicable to risk management and forecasting, where the ability to identify intricate patterns is crucial to making predictions. Nearly 100 companies reportedly are experimenting with neural network development.

Where do neural networks fit in?

In theory, a neural network can be put to work in any application where substantial amounts of data are used to predict outcome. Neural networks are being used today in applications ranging from analyzing engines to finding submarines.

For instance, one credit card company is using a neural network to identify the fraudulent use of plastic by spotting unusual purchasing activity. Neural networks are also being employed--with profitable results--in securities and options trading.

One insurance company is experimenting with a neural network to compare an individual agent's handwriting stored in a personal computer with the handwritten policy application forms the agent produces. Through pattern recognition, the computer can read each agent's handwriting, putting accurate data into memory without going through the potentially costly, error-prone method of having a clerk enter the data through a keyboard.

Another system, this one under development, will permit banks to use a neural network to recognize handwritten numerals on checks. Hundreds of thousands of checks will no longer have to be entered into the system by processing clerks.

Virtually all computers today operate through linear programming--applying complex sets of rules and thousands of yes-or-no, what-if answers to produce output. The fascination with and the practical significance in neural networks is that they make up their own rules. The more decisions they make, it appears, the better those decisions are.

But understanding how neural networks work is a job for Superman, a Ph.D. in mathematics, and perhaps both. Yet any CFO, and even CEO, risks falling behind in the competitive race without having at least a rudimentary knowledge of what neural networks are and just how they might be applied, so then he can decide whether or not to implement such a system.

Rather than explore theory, look at a practical application that we're using at Sears Mortgage Corporation. It's the introduction of a neural network system to make the same kinds of decisions now being made by mortgage underwriters. The decision to be made: will a mortgage be approved or denied?

Before looking at how a neural network can help answer that question, let's put the system in context by outlining its various bottom-line advantages. These include:

* Growth--Sears Mortgage, now one of the 10 largest mortgage originators in the U.S., expects to more than double its customer loan base within the next five years.

The average underwriter can make a good loan decision in about 45 minutes. The neural network can make that same decision in a second. Neural networks, we believe, will enable us to double originations while maintaining the underwriting staff at its present level. * Better decision-making--Good lending decisions keep foreclosure rates low and bring the maximum number of sound loans into the portfolio. Poor decisions can increase the foreclosure rate, inhibit the ability to sell loans to the secondary market, or result in unnecessary loan turndowns, in effect turning away profitable business.

The neural network will process the 60 to 70 percent of applications that represent clear-cut, easy decisions at both ends of the spectrum. Underwriters will then be able to spend more time evaluating the tough applications in the middle that require the greatest level of expertise.

Rather than outright rejecting some applications, underwriters will have more time to review problem situations and reduce lost business by developing alternative financing packages that can then be accepted. * Consistency--Unlike human decision-makers, neural networks never have a bad day and they never get tired. A neural network will, therefore, improve the overall quality of underwriting decisions.

Underwriters base decisions on the sum total of their individual experience. Neural network decisions are based on the collective data representing a large number of lending decisions made in recent years by every underwriter on the company's staff. In the future, they also will consider every new loan that has been accepted and evaluate the loan performance. * Alleviation of seasonal peaks --Heavy influxes of mortgage applications during the spring and summer months will be handled easily without resorting to excessive overtime or adding temporary staff. * Improved consumer service --Processing mortgage applications faster and reducing the workload of underwriters, who may review fewer applications but spend more time on them, we believe could cut the average processing time by a third. The system is also expected to lower the turndown rate by as much as 4 percent with only a 2-percent increase in risk. As for the cost of setting up a neural network, since the systems are so customized, the range of investment dollars is wide. Networks that run on a personal computer, for instance, could cost a company as little as $50,000, or even less. And that figure takes into account the entire package, including software development. On the other hand, if a network is sufficiently complex to require implementation on a mainframe, the investment could grow to $500,000 or more. In short, the size of the project dictates your expense.

But how does it work?

The neural network develops the ability to decide and then learns to improve through massive trial-and-error decision-making. The initial phase of the development of the system is called supervised learning.

A network is "trained" in this instance by the company supplying key data and specifying the correct outcome--according to what an experienced underwriter decided when he or she reviewed the same data.

Mortgage lending decisions made by underwriters or a computer are based on four main factors: the adequacy of the borrower's income to pay the monthly mortgage, real estate taxes, and insurance while meeting financial obligations; the stability of employment; the credit history of repaying debt; and the value of the property as collateral.

Underwriting decisions are subjective. There are few hard and fast rules. Most people's mortgage credit worthiness falls within the middle range of a spectrum. That's why a neural network that takes into account the interrelationship of a large number of variables is the best, though still imperfect, substitute for human decision-making.

The neural network system for mortgage approvals, developed in a cooperative effort with Nestor, Inc., a specialist in what is called neural net technology, utilizes about 30 sets of data. All of the data for each applicant has a direct bearing on the decision and interrelates in complex ways.

Data ranges from the applicant's zip code, the age of the home, and its appraised value to the monthly income, proposed housing expense, and credit rating of the borrowers. The computer's job is to make hundreds of thousands of calculations for each application, searching for patterns. How does the data in one application compare to the pattern of data for all approved loans stored in memory, and how does it compare with denied loans?

Initial neural network decision-making in our system is based on 5,000 real mortgage lending decisions--2,500 acceptances, 2,500 rejections--made over the last two years by expert underwriters.

For simplification, picture a bottom row, or layer, of small circles each representing a decision-making variable and each containing a formula that produces a financial ratio. At the beginning of supervised learning, each formula is arbitrarily weighted according to its perceived importance in determining credit acceptance or denial.

A second, middle layer of circles, called the hidden layer by neural network developers, is designed to approximate how data is collected and assimilated by the brain. This secondary layer is essentially a map that records how each variable is related to every other and how it should react to new data.

Every bottom circle is interconnected with every other circle through this hidden layer much as the brain consists of vast networks of synapses between various points.

The top layer is the intended result. This series of circles in this simplified overview stores data on correct outcomes, whether underwriters approved or disapproved various loans put in memory.

Learning from experience

In the supervised learning mode, the network begins to compare all variables with all others in every possible combination in an attempt to achieve the correct final result.

Expressed in terms of a number, again a simplistic way of looking at it, an approved loan might be equal to one. Making calculations at blinding speed, the network in effect keeps improving the accuracy of its weighting system for each variable in an effort to bring the answer for any given loan as close as possible to one.

The weighting of variables is continuously modified to bring each loan to the proper approved or denied status that was determined by the underwriter who originally evaluated the loan. Proximity to one in this example would mean approval; to zero, denial.

After sifting through 1,000 approved and denied loans, for example, the network achieves one level of accuracy as it constantly changes the weights to identify patterns and find the best combination. With each succeeding loan imputed, through continuing trial and error, it improves its level of accuracy. The more applications it reviews, the more it learns and the more accurate are the decisions.

After the network has absorbed the information on 5,000 previously approved or rejected applications, it is ready to begin processing new loans. The present staff of underwriters reviews the network decisions for some months to ensure that correct decisions are being made.

Application loan data is fed into a central mainframe computer from local branches. Approval or denial decisions can then be accessed from an ordinary PC, which will show a decision or indicate it has not made one when the data is inconclusive.

The computer assigns a confidence level of from 0 to 100 percent to each application processed. For example, when the network processes what we like to call a "slam-dunk" loan--one with a monthly income more than adequate, plus a credit rating and all other parameters fitting into a highly acceptable pattern--it will peg its accuracy comfort level at 100 percent. An inconclusive set of data will yield a low comfort level percent.

The comfort level that has to be achieved before the network will make a definite recommendation is programmed into the software. The computer also will identify the reason for turndown recommendations. Most turndown decisions will be routed to underwriters for a traditional "hands-on" review.

Beauties of the system

One of the special beauties of the system is the neural network's ability to respond even when it is receiving what we call fuzzy data. Fuzzy data is incomplete information or data so far off pattern that the computer has a tough time making choices. Yet the computer will still try to deal with unfamiliar patterns it finds, developing a best approximation of what the result might be.

Another amazing attribute of neural networks is their ability to spot patterns in decision-making that old-fashioned, rules-based linear systems of evaluating mortgage decisions would never pick up.

In adding or subtracting points for each variable, for example, a rules-based linear system might penalize a member of the Armed Forces for changing residences frequently. A neural network, like a real-life underwriter, would have the capability to recognize the aberration, penalizing the typical applicant but ignoring the frequency-of-move rate for military people.

The network would have learned that moving frequency is not a measure of instability with military people as it is with the general population. Neural networks don't have to be "told" this--they have the power to discover such subtleties. In effect, they are so good at identifying patterns that they even have the potential to outsmart underwriters, who may not be aware of certain peculiarities on applications.

They're here to stay

The next step forward for neural networks is expected to involve new computer hardware. So far, neural networks use conventional computers that employ serial processing technology. The first generation of new parallel process computers now under development may make neural systems an even more effective decision-maker.

So, in the final analysis, what is the bottom line? Will neural networks become a permanent part of corporations' decision-making methodology? Or are they a rapidly approaching shooting star that will shine brightly only to vanish when another promising technology comes along? If a neural network had been trained to give a yes or no answer on its own future, it probably would not as yet see a discernible pattern. It would be looking for more data.

Being on the leading edge of technology has its risks, yet also tremendous rewards. We at Sears Mortgage Corporation want to leverage technology to provide superior customer service and to gain a competitive advantage. Although neural networks continue to struggle through their infancy, from where we sit, it looks like they have a very good chance of growing up.

Mr. Smith is a vice chairman of FEI's Committee on Information Management.