Neural networks: the buck stops here.

Neural networks: The buck stops here

Neurocomputers are a breed of rapidly developing hardware on which artificial neural networks are trained to solve problems. Because these systems sort through immense amounts of information and pick out patterns from the onslaught of data, they may become useful tools in complex financial decision-making, according to computer scientists who presented reports in San Diego last week at the IEEE International Conference on Neural Networks.

One such neurocomputer-driven neural network, developed by Edward Collins and his colleagues at Nestor, Inc., in Providence, R.I., accurately makes decisions on mortgage risks commonly evaluated by mortgage underwriters.

Mortgages are usually underwritten by both a mortgage provider and a mortgage insurer, says Collins. A variety of information is considered before a mortgage is granted or denied, and disagreement is not uncommon between provider and insurer.

The researches designed a neural network with three internal layers of processing elements. The strengths of connections that transmit messages between elements are altered as the system trains itself to achieve a desired output. In this case, the network was given information from 5,000 mortgage applicants; decisions on the applications made by a mortgage underwriter served as a training signal. Data fed into the computer covered each applicant's background and financial history, as well as the type of mortgage required and the property being sought.

Each layer of the network analyzed a piece of the complex financial input and determined the riskiness of granting a loan. A "controller" built into the neurocomputer then determined whether there was significant agreement between the three layers and, if agreement was reached, rendered a response.

When the statistical rules followed by the controller allowed for agreement in each case, the resulting decision agreed with that of the mortgage underwriter 82 percent of the time. When the statistical criteria for agreement between the three layers were tightened, says Collins, a response was obtained for one-third of the cases with 96 percent agreement.

Furthermore, notes Collins, the neural network was better than the mortgage writer at predicting who was a good loan risk and who would default. The system's three decision-making layers appear to enhance human judgments on loan applications, he maintains.

A similar neural networks, designed by Shashi Shekhar of the University of California at Berkeley and a colleague, trains itself to rate the quality of bonds purchased by investors. Ratings reflect the probability of making a profit on a bond from a particular company. Financial information on a company is evaluated by ratings authorities who use standard mathematical equations to aid in their decisions.

When fed detailed, publicly available financial information on companies issuing bonds, the neural network predicted established bond ratings better than the typical mathematical procedures used by bond raters, says Shekhar.

"Neural networks provide a more general framework for connecting financial input about a company to an output, the bond rating," he asserts.

The stock market, however, is a tougher nut to crack. Economist Halbert White of the University of California at San Diego recently provided a neural network with daily rates of return on IBM stock over 500 days in the mid-1970s. The network did its best to extract predictable fluctuations in the stock's worth, White says, but so far only random jumps and dips are evident.

"It won't be easy to uncover predictable stock market fluctuations with neural networks," he remarks, "and if you succeed, you'll probably want to keep it secret."