Artificial Intelligence: a view from 1990

Spread the love

If neural networks are such great pattern matchers and can be used for prediction and forecasting, then can they be used to predict the stock market? If so, we can all get rich. Naturally, such thinking is to be expected and someone tried to do it [White, 1988]. He used NLS and feedforward neural networks to predict daily IBM stock prices. He also used back-propagation for training. Unfortunately, the results were disappointing. In some ways, this result could have been expected. After all, neural networks can only process information, make data transformations and detect patterns. They cannot make up something from nothing. Where no information exists, neural networks cannot magically find meaning. Assuming that stock prices are nearly random on a day-to-day basis, then neural networks cannot be expected to predict the next day’s stock price. However, negative results do not prove that the task of predicting stock prices   cannot be done.
More realistically, neural networks may prove to be useful if more reasonable problems are worked on. For example, HNC used neural networks to analyze foreign currency trading. Their system was able to discover features in the data. They analyzed information about the Pound Sterling, Japanese Yen and Deutsch Mark. With this neural network system inexperienced traders could make profitable decisions. Others found that feedforward neural networks were substantially better than regression techniques when used for corporate bond rating [Dutta & Shekhar, 1988]. Using their neural network systems, Nestor Company developed a successful automated securities trading program. They correctly classified 75 percent of the patterns which were prescreened for unambiguously identified patterns. This result was good when compared to other automated traders which operate in the range of 50 percent to 60 percent accuracy. Their system was done on a DEC VAX and processed a pattern in about one second. A bond-trading system was developed by the Nestor Company which could make correct recommendations 72 percent of the time. A non-neural network system only gave correct signals 55 percent of the time. While the percent improvements were small, the profitability was significant.
It appears that neural networks can be used to an advantage for trading or market tasks. Possibly one of the secrets to successful market analysis, is to define and restrict the problem to one which is potentially solvable. Neural networks can not identify information which is not available in the original data. However, if there is information in the raw input, even if it is hidden, neural networks seem to be able to pull it out. The successful examples above suggest that neural networks can be used for profitable trading. Their success suggests that more consideration should be given to using neural networks for financial analysis.

Alianna J. Maren, Craig T. Harston, Robert M. Pap, HANDBOOK OF NEURAL COMPUTING APPLICATIONS, Academic Press, Inc., San Diego, California USA, 1990

ISBN 0-12-546090-2
I love doing things that are considered (almost) impossible 😉