VantagePoint : The History and Controversy Behind the Neural Network Theory

Company News and Press Releases

‘SMART’ systems

July/August 1997 By Alison Kahler MANAGING EDITOR

Neural network theory applies the principles of artificial intelligence to the analysis of financial markets. Alison Kahler explains the history and controversy behind the process.

It seems the obvious solution to the quest for accuracy in trading decisions. Why not combine the power of the human mind with sophisticated computer technology to build a new breed of trading system? Artificial intelligence was ordained the ‘next big thing’ when it was first applied to the analysis of United States financial markets in the early 1980s. ‘Neural network theory’ was applied at a feverish pace to technical analysis of markets. Perhaps fooled by the term ‘intelligence’, some commentators and traders built up a whirlwind of hype around neural networks. But initial hiccups in implementation and criticisms of ‘over-optimisation’ brought about a lull in interest in the new systems. Once the dust and unrealistic expectations of a ‘Holy Grail’ settled, the practical uses of artificial intelligence became evident. It now seems certain that the process is more than a passing fad or simply an overly-scientific form of technical analysis. Reports from the United States estimate that thousands of companies use neural network software to mathematically model a plethora of tasks. Locally, awareness and interest in the feasibility of artificial intelligence is said to be growing. A recent University of Melbourne seminar focusing on neural networks attracted personnel from Tullamarine airport seeking to forecast flight demands, as well as participants looking to predict electricity pricing. Earlier this year, an organisation in Melbourne was reported to have spent $20,000 commissioning a presentation outlining the pros and cons of neural network theory. In terms of financial markets, the US-based Futures magazine is a guide to industry acceptance of the principles of artificial intelligence. Early articles focused on the purported ability of neural networks to analyse data across markets. After a gap in reporting, the next wave of articles focused on the deemed over-complexity of the process, along with a myriad of problems in practical implementation. More lately, the emphasis has been more along the lines of “this is valid … just don’t get carried away or think that it easy”.BRAIN POWER Otherwise known as ‘artificial neural systems’, ‘adaptive systems’, ‘neuro computers’ and ‘intelligent systems’, neural networks are a bit like the ultimate form of database. In simple terms, a properly programmed system takes in both technical and fundamental factors, analyses the data and then makes trading decisions of its own accord. The nature of this proposition is understandably attractive to financial market participants. Getting a little more complex, neural networks are different to conventional analytical methods because they do not employ pre-defined trading rules in the search for market signals. Instead, through what is called an iterative training process, neural systems learn the underlying associations and casual relationships within the technical and fundamental data impacting on the price of a specific equity, commodity or a stock index. Learning is possible because the network is taught to make generalisations and change emphasis when told the generalisations are wrong or right. Once a system is trained, it is used to predict subsequent prices and the future trend direction for a given market. The term ‘artificial intelligence’ is used to describe such systems because, in essence, programmers are building a computer model that mimics the way a collection of brain cells – or ‘neurons’ – operate. By copying human thought processes, system designers are seeking to build a model that has the ability to learn from experience, to develop rules and recognise patterns in data.THE NEURAL ADVANTAGE Proponents of artificial intelligence claim it has several advantages over traditional forms of technical analysis. “I believe the application of neural networks to intermarket analysis will play a major role in broadening the scope of technical analysis,” says Lou Mendelsohn, a US-based systems developer who has been involved in artificial intelligence since it was first applied to trading. “Today’s financial markets are interdependent. Understanding their influence on each other is critical for any trader but traditional analytical methods tend to focus on individual markets. “Single market analysis is too restrictive, yet still so prevalent among novice traders. It’s no wonder the loss statistics are so high.” Mendelsohn claims that the pattern recognition capabilities of neural networks are near to a revolution in trading systems. “Traditional barriers between technical and fundamental factors are broken down,” he argues, “creating a broader analytical framework.” Mendelsohn terms this framework ‘synergistic market analysis’ because it helps discern hidden patterns and relationships between seemingly disparate data. The fact that artificial intelligence allows ‘non-linear’ measurement is touted as a significant plus of neural networks. Most mathematics predicts measurable, consistent – or linear – change. Non-linear dynamics apply to volatile or inconsistent change and are, thus, seem by some as the best means of analysing market behaviour and predicting price volatility. Critics of neural network theory claim that it is the ultimate curve-fitting exercise. The term ‘curve-fitting’ means that traders add rules to their system until it achieves a perfect hypothetical performance based on historical results. As history never repeats, curve-fitted systems generally fail when applied to real trading.NO SUBSTITUTE Neural networks are just one branch of artificial intelligence. Knowledge-based systems and ‘fuzzy logic’ also fall into the same realm, with the latter becoming increasingly popular among traders. Genetic algorithms are an offshoot of artificial intelligence technology. The personal computer revolution made all of the above not only accessible but also affordable for a much larger proportion of the population than in previous years (studies of artificial intelligence date back to the 1940s.) In Australia, professional users of neural networks, at least in financial markets, seem to lie low. One funds management organisation in Melbourne uses a neural network to trade the Share Price Index. Despite the ongoing development of the industry, neural nets remain the most well-known application of artificial intelligence to financial markets. This is probably because traders have been seduced by the concept of ‘smart systems’ with the ability to ‘learn’, an attribute more commonly applied to neural nets than the other forms of the technology. But neural networks can’t think on behalf of a trader. “Artificial intelligence doesn’t have anything to do with a user not having a knowledge of market behaviour,” says Lou Mendelsohn. It seems fairly obvious but maybe it needs saying – the best systems developers are those who understand market dynamics. This is true whatever technology is utilised in the development of a trading system.