Mr. James O'Sullivan, of O'Sullivan Brothers Investments, Ltd. (Connecticut) has been successfully using many BrainMaker (California Scientific) neural networks on a daily basis for three years to do financial forecasting. He earned $250,000 in one month using neural networks to advise him on his New York Stock Exchange seat trades. Some of his networks are 88-90% accurate in their predictions, according to Mr. O'Sullivan. He uses an automated neural network system that monitors more than twenty different financial markets on a daily basis.
Mr. O'Sullivan has some unusual designs which act more as detectors of specific market conditions, rather than as exact price predictors. He combines the neural network data with other data from his technical analysis software to produce an automated report about a certain market. He gets his data live via satellite from Data Broadcasting and puts it into a charting and technical data module. He has pre-programmed the various algebraic manipulations to be performed on his data before BrainMaker files are built. He does moving averages, changes from period to period, and a few proprietary operations. He runs new data through the system and produces the one-page report in about thirty seconds. He says at least 80% of his decision-making is based on neural network predictions.
Mr. O'Sullivan has not fully disclosed his neural network designs to us, but his basic insights are still quite valuable. The key is to ask the neural network the right kind of questions. He asks questions such as "What is the probability of the product (or market) going up 0.618 standard deviations?" and "What percentage of the time does it go up that much?" He also asks questions about the directional behavior of the market and at what price the product (or market) is likely to take off in one direction or the other.
Mr. O'Sullivan's neural networks output several different things such as predicted prices, limits, and directional thrust. One neural network outputs the probability of a certain price occurring the next time period. Another neural network produces best stop price and best target price for long and short positions. Other neural networks produce directional indicators for three market energies. Another predicts the level at which the market is likely to take off.
In one design, the network is given various market conditions as input. The training output is the likelihood of various changes in price. For example, his neural network is told during training that, given similar market conditions, the closing price goes up at least 0.313 standard deviations above the prior day 90% of the time, at least 0.618 deviations 80% of the time, and at least 1 deviation 70% of the time.
An interesting phenomenon of the market is that when a change starts occurring in one direction or the other, there is a point at which it is very likely to continue moving in that direction for several time periods. Once a price reaches that level, there is a reduced risk to buy or sell (whichever is appropriate). Mr. O'Sullivan calls the network that predicts this price level his Risk Barometer network. He uses all the neural networks trained for a specific market when making decisions. For example, if the long term trend is up, the Fast Movement network is a large positive number, his Risk Barometer says 233.092, and the NYSE is at 250, it could indicate an overreacted market that will reverse itself soon.
|For Trades: Long buy Short sell|
|best stop = 230.708 231.992|
|best target = 232.054 230.646|
|Trend||Reaction||Fast Movement||Risk Barometer|
James W. O'Sullivan, Neural Nets: A Practical Primer, AI In Finance, Spring, 1994.