What are the problems then with the use of neural networks in finance ? Well, to understand this one needs to understand the implications and intent of using a neural network, what can they predict and what can’t they predict. First of all, the idea of a neural network is to predict a given result with a previous “training” on a data set of the same characteristics as the data set in which the neural network would be used. For example, if you want to predict the EUR/USD price you would first train the neural network over the past years of EUR/USD price data. After this training you can then try to make a prediction based on the adaptations of the neural network to the previous set.

The problem with this approach is the basic premise that says that there is always a function to fit a given set of parameters. Imagine this example, try to tell me what the next number in the series would be : 1,2,3. You say 4 ? No, its 5. What is the next number (1,2,3,5) ? 8 ? No. It is 50. This means that no matter how complicated a data series is, you will always be able to find some mathematical expression that predicts all the series and then gives you the next number. Since there are infinite possible mathematical expressions to describe a data set, it is impossible to know if the mathematical expression one has has any resemblance to the actual function that determines the next result on the data set.

For this reason, absolute variables that come from a very complicated combination of factors, which include random factors, are not at all predictable using neural networks. Price, for example, is very badly predicted by neural networks because price depends on many factors, some of them being random. What can a neural network predict better ? Actually, neural networks are good at predicting information which comes out of cyclical aspects of the market. For example, the prediction of overbought and oversold levels, the end of market cycles, the beginning of a down/up trend are all things that neural networks can predict with better accuracy. Variables that determine risk, such as the probability of a position of a given system being profitable or non profitable is also predicted to a good extent by neural networks.

In conclusion, neural networks are not good by themselves as systems because they are not very good at determining price levels, entries or exits. However neural networks have merit as “filters” in that they are able to predict the risk of opening positions with much better accuracy. Of course, the extent to which this accuracy is true depends on the actual characteristics of the neural network and the characteristic of the system it is filtering. None the less, it is safe to say that neural networks are not a holy grail and they have as much merit as any other trading strategy or filtering strategy. As a matter of fact, neural networks are often a much more complicated and undesirable approach to trading strategy optimization, usually the use of genetic algorithms (probably a post will be out on this next month), which allow us to see the mathematically optimized expressions is preferred over the use of neural networks.

If you would like to learn more about automated trading systems and how they are optimized and programmed to give long term profitable results please consider buying my ebook on automated trading or subscribing to my weekly newsletter to receive updates and check the live and demo accounts I am running with several expert advisors. I hope you enjoyed the article !