Neural Networks in Trading: Why One Pair and Not Another?

A large majority of new traders will be very eager to tell you that if you have something that “works” then it should work on all currency pairs, if you really have some inefficiency which is truly fundamental then this will work on the EUR/USD, the USD/JPY and the meat and soy futures, if your understanding is fundamental then your inefficiency should prevail amongst all or at least most market instruments. However when we start to study the success of Neural Network strategies we find out that they are not always successful in all pairs and their results can change drastically from one pair to another. Why is it that a neural network can succeed greatly in one pair and fail miserably on another? What makes a neural network successful?

On today’s post I will talk about the issue of neural network success and why my experiments with this tool lead me to believe what has been evident through the development of trading systems using manual methods and genetics, currency pairs are very different and respond to very different phenomena. Through this post I will share with you some of my results using neural networks and why I believe the question of profitability is largely relevant to liquidity (as I have mentioned on past posts) and the information used to interpret market data.

When I first started to experiment with neural networks the first thing I found which worked dealt with a simple usage of currency pair data to predict changes in daily closing direction. This system – which became Sunqu – is currently still under development but has shown very profitable results on the EUR/USD through a method which involves continuous adapting of the system against the past 100 days of market data every 5 days. Now the interesting thing is that you would think that such a system would work on other pairs as well (the GBP/USD for example) but you find that in fact the system fails miserably on almost all pairs except the EUR/USD and the USD/CHF. Why is this the case? Why does a system that has such highly adaptive profiting power on the EUR/USD fail so badly on other pairs?

The answer seems to be related with the relevant information to the different instruments. On the EUR/USD the market seems to give us a large amount of information which is enough to actually come up with exploitable inefficiencies while this same information seems to be “lacking” in the case of other pairs. For example on the USD/CHF similarly profitable results are only achieved when we increase the learning period to 200 days showing that there is a need for much more that than on the EUR/USD. On other pairs the problem is even more pronounced coming to a point where we have a simply abrupt failure of the system to work, the information the neural network has is simply not enough to predict outcomes with an edge on these pairs.

Why is the GBP/USD so fundamentally different from the EUR/USD? Isn’t an adaptive system supposed to work on all of them? My research right now points to the fact that information needs to be weight in different ways for the different currency pairs as they react differently to certain kinds of data. For example the GBP/USD doesn’t contain enough information in its own data to predict outcomes but if you give the neural network Dow information you start to get somewhere. The information contained in the Dow is a leading indicator to certain outcomes in the GBP/USD while it isn’t that relevant for the EUR/USD at all.

Doing a principal component analysis (PCA) to find out what sources of information might be correlated also reveals that using information from other pairs is not at all relevant. For example the EUR/USD system doesn’t have any increase in predictive power when using USD/CHF or GBP/USD data and the same is true for these two currency pairs. It is interesting to note how neural networks seem to react naturally to whichever association are generally considered fundamental. For example a successful system for the USD/CAD can be built if you not only consider its own data but also the Gold and Oil prices which seem to hold a strong relationship with this pair. Stocks and interest rates are also very interesting data to achieve predictions for carry trade systems.

The world of neural networks has therefore confirmed several of the suspicions I already had about currency pairs. The easiest pair to trade is the EUR/USD (as the information within this pair is already very relevant to predict its own outcomes) while other pairs are much harder to predict since their outcomes are the result of more complex interactions dealing mainly with macro-economical factors. This doesn’t mean that the EUR/USD isn’t influenced by these factors as well but it means that you can be profitable by simply trading the EUR/USD with EUR/USD data while this is harder to achieve for other pairs. Of course I am still a “youngling” in this Neural Network game and time will certainly reveal deeper and probably much more fascinating truths about the market :o)

If you would like to learn more about my work in automated trading and how you too can learn how to code, build and evaluate your own systems please consider joining, a website filled with educational videos, trading systems, development and a sound, honest and transparent approach towards automated trading in general . I hope you enjoyed this article ! :o)

Print Friendly, PDF & Email
You can leave a response, or trackback from your own site.

One Response to “Neural Networks in Trading: Why One Pair and Not Another?”

  1. xyz says:

    Excellent post. The idea of testing whether the information given by the price movements of a given pair is enough to predict the future movements on this pair is a great insight. A safe option is to restrict the trading to the most liquid pairs (EUR/USD, USD/JPY, EUR/JPY, with perhaps a few others).

Leave a Reply

WordPress › Error

The site is experiencing technical difficulties.