The Application of Neural Networks to Credit Scoring.

Of the credit scoring methods currently in general use, neural networks is the least well known. However, it may be the most powerful decision support technology available for solving this problem. This article will introduce you to neural networks and take you through a typical step-by-step process for developing a credit scoring system based on neural network technology.

What are Neural Networks?

For those of you unfamiliar with this technology, I will briefly describe what neural networks are and how they differ from conventional computers. While neural networks will not replace conventional computers, when they are used to augment them, the result can be a system that not only rapidly crunches numbers but also possesses human-like qualities. People are not well suited to perform a large number of numerical calculations accurately. Computers accomplish this job superlatively. On the other hand, computers do a poor job of recognizing a person's face--something that humans do very well. A neural network, patterned after the structure of the human brain, can perform certain human-type tasks and is a very valuable tool when applied to the proper problem, such as credit scoring, which essentially is a synthesis of available information extrapolated to produce an estimate of the risk inherent in future performance. When neural networks are applied to the problem, the available information is seen as a patt ern that can be categorized to estimate future risk based upon the data that was used to train the system.

Introduction

Conventional computing is founded on underlying principles of logic and mathematics. The architecture used by most computers assumes a central processing unit connected to an area of memory that contains a stored program, executed in a sequential manner by the processor. Neural computers take an alternative approach in that they directly model the biological structure of the brain and the way it processes information. A new kind of computer architecture is necessary that, like the brain, consists of a large number of interconnected processing elements (neurons), which operate in parallel.

A neural network, therefore, is an example of specialized parallel-processing architecture. Neural computing should not be viewed as a competitor to conventional computing, but rather as a complementary technique. So far, the most successful neural computing applications have operated in conjunction with other computing techniques. For example, a neural network can be used to initially analyze incoming data, and to classify that data into various categories, each of which may require different processing. The resulting stratified data is then passed to a conventional computer system where the remaining processing is accomplished.

Neural networks can be taught to perform complex tasks and do not require programming, as do conventional computers. They can learn from experience, they can generalize from examples, and they can extract essential characteristics from noisy data. They require significantly less development time and can respond to unspecified situations or to situations that were not even contemplated during the networks' development.