OASIS Financial

What is an artificial neural net?
Neural networks are the result of scientists and the engineers attempting to exploit the process used by the most complex computing device known to man - the human brain. The brain's powerful thinking, remembering, and problem-solving capabilities result from its ability to use a complex web of tens of billions of neurons. These neurons allow the brain to process, analyze and remember the massive amount of data that it is exposed to every day through the specialization of the neurons. The specialization allows the brain to break complex problems into small parts and allow the neuron to concentrate on their individual part of the problem and then share their answer with the rest of the brain through the interconnections. An individual biological neuron has fairly simple computational ability. By itself, an individual neuron is not very interesting and has limited computation ability. The interesting computational properties emerge when neurons are combined together in various ways. A simplified diagram of one type of neuron is shown below:

Artificial neural nets exploit the concept of densely connected networks of simple processing units for practical computational use. Neural nets have been successfully developed in areas of credit card fraud detection, stock market timing systems, consumer credit risk analysis, bankruptcy prediction, and military weapons systems. Neural computing in general, builds models based on historical data. Neural networks are applicable in any situation where there is an unknown relationship between a set of input factors and an outcome, and for which a representative set of historical examples of this problem are available. The objective of building a model is to find a formula or program that facilitates predicting the outcome from the input factors.

The Neural Credit Assistant is a good example of how a neural network operates. The goal of the model is to develop an algorithm, or process by which to determine the senior rating of a company through an analysis of financial ratios derived from the company's financial statements and senior rating assigned by rating agencies. Our training examples are all contained in a proprietary database of the companies' financial information and senior debt financial information and senior debt ratings. A company typically has multiple examples within the database, one example for each four quarters of data. The companies selected for the system were those with ratings based upon their own financial position and not dependent upon credit support from a parent company or other entities. The financial data for each example includes historical data going back three to five years from the balance sheets, income statements and the equity markets. The bond rating used for each example was the rating two quarters following the end of the period. We used the rating two quarters after the end of the period to adjust for the time lag between the end of a period and when the financial information for that period is actually reported. If the rating agencies are going to adjust the rating based on the financial results, then there will be a lag between the end of the period and a rating change. The database is continuously being added to as new companies are being rated and as new financial statements are being released. The Neural Credit Assistant currently uses over 40.000 examples from over 3,000 companies in its learning process. Each month the Neural Credit Assistant is retrained to take into account new examples from the current companies, rating changes and examples from new companies.