There is a long road ahead for neural networks to be able to achieve consistent gains and to be democratized, but the path is royal.

IT consultant in London and author of a website dedicated to trading, Nicolas Vitale explains that neural networks are part of modern techniques developped in recent years in trading, where some conventional methods, such as indicators, or moving averages were becoming a bit outdated. Before the widespread use of computers, one had to do things by hand. moving averages were then a relatively modern technique, but now everyone knows how it works, everyone would buy basically the same thing at the same time. Therefore, there is no more advantages to use them. This does not mean that it never works, because in the long term and with flair, moving averages, in addition to other techniques, can be effective.
"Still, there are more modern techniques today that include neural networks or genetic algorithms. "

How does actually work a neural network? It is based on a neural structure with a network and data (unemployment rate, changes in bond prices, inflation, stock prices, etc. ...). Each neuron receives as "input" these data which act as stimuli, on the same principle as the biological neurons. These neurons weight each data and send a signal if the stimuli exceeds a given threshold. The weight’s coefficient is then adjusted according to the "output", that is to say, decisions to buy or sell, and so on, with a stronger factor for relevant information, to expense of those which are not. The neural network sorts itself and sets its own parameters based on the data tested.

The difference and confusion sometimes made with genetic algorithms is not simple to understand, especially since both are often associated. After all, genetic algorithms like neural networks are also based on evolution theory, neural networks aiming to find the best solutions to "survive on their own." "Genetic algorithms," says Nicolas Vitale, "can not go through brute force. We start from a more or less random strategy, and parameters evolve. The parameters that produce good results survive and reproduce. The others disappear. It is really Evolutionism applied to the algorithms. "

There is lot of improvements ahead for these extremely complex systems to become widespread, and reach the same level of automaticity and efficiency of "traditional" algorithms. "A neural network cannot survive and be efficient without important analysis work and monitoring. The "machine" is not autonomous. If you run the program and decide not to intervene, you will inevitably loose money at one time or another. There is always a lot of research works to do. To end up with a real predictive system, one have to test and optimize in the past, test again on hidden data, and do it again and again. Lot of people talk about neural networks these days, but that does not mean that everything is automated, and that one can go home to sip a cocktail without real time monitoring. "

However, systemic risks are reduced since these types of algorithm are mainly used by hedge funds, for now anyway.

Regarding the different types of neural networks, forecasts and trading experiments on forex indicate that Higher Order Neural Networks (HONN) and Multilayer Perceptron (MLP), coupled with various statistical techniques or merely techniques, outperform other types of neural networks such as recurrent neural network.

According to Christian Dunis, director of Centre for International Banking, Economics and Finance, "HONN and the MLP achieve robust performances and exceed consistently, in simulations, all the other models. Even when transaction costs and leverage are applied, these two types of networks are doing better than other neural networks or statistical models in terms of annualized returns, and considering all the windows studied.