Neural Networks for Not Yet Techies

Neural networks are systems of programs that attempt to duplicate the way the brain works, as a huge system of interconnected neurons. It is a branch of Artificial Intelligence / AI. The original neural networks used chains of actual circuits, but now they are really SNNs -- Simulated Neural Networks. The computers use the neural network algorithm.

1. ADALINE
2. Backpropagation
3. Self-organizing
4. Bidirectional Associative Memory (BAM)

Neural networks are a set of interconnected nodes that accept input, process it and output it to another step in the process. It involves a large number of processors operating in parallel, each with its own knowledge and data access in local memory. They are initially "trained" with a large amount of data and rules.

Neural networks became popular with financial traders in the early nineties -- or at least with techies who hoped to use their geeky knowledge to make a killing in the financial markets.

They gave their neural networks an enormous amount of market data relating to stocks, bonds and commodities and their historical prices over many years. They believed that they could train the networks to learn the patterns in price action that would reliably signal future prices moves, up or down.

If a financial stock or commodities trader knew a price pattern that would reliably predict a significant price move in a market, that would be a ticket to making a fortune.

I suspect that a lot of traders are still trying to train their neural networks to learn a price pattern that will make them rich. Unfortunately, I don't believe it's possible because reality will always be more complicated than a computer model. For instance, in August 1991 how could anybody (except the terrorists, who did use their "inside" knowledge of what they planned to make money in the markets) have known that the airline and travel industry was about to face a tremendous disaster? September 11 is not the type of event that can be "discounted" by markets before they happen.

• gradient-based training
• fuzzy logic
• genetic algorithms
• Bayesian methods

In feedforward systems, learned relationships about data can be "fed forward" to higher levels of knowledge. They can also learn temporal concepts and are widely used in signal processing and time series analysis. They are appropriate where the data they learn does have clear cause-effect relationships.

• oil exploration data analysis
• weather prediction
• the interpretation of nucleotide sequences in biology laboratories
• exploring models of thinking and consciousness

Neural networks might be helping to make some traders rich. Although I believe that a lot of market price action is random (and therefore unpredictable even by supercomputers), there are no doubt some news-driven and fundamental business reasons for some short term price rises or falls, and so they might help you to find such patterns.

Next: PCI Peripheral Component Interconnect