The Challenge of Forex Trading for Machine Learning

3 min read

Machine learning is a branch of artificial intelligence that has grabbed a lot of headlines previously. People are fascinated by the concept of machines seemingly ‘thinking’, and learning how to carry out tasks more proficiently over time. This has been both a holy grail of computer programmers, and a mortal fear of the human race. But machine learning is no longer something over the horizon; it’s here right now. 

Cut through the paranoia and perturbation, though, and you soon find that machine learning has a ton of useful everyday applications. And one area where machine learning can potentially help out is in the field of trading. AI is already outperforming humans in some real-world tasks, so can machine learning hope to dominate the multi-trillion dollar market that is Forex? 

Building machine learning strategies and techniques that enable machines to learn in real time, and thus deliver in market conditions, is pretty much the exalted goal of algorithmic trading. The notion of a computer that can deliver Forex results time and time again is obviously an attractive concept. However, it has proved difficult to achieve this as of yet. There have been some promising dawns, but no academic work has managed to demonstrate a machine that can consistently outperform humans in this field.

Limited Fields and Determinism

The problem that machines encounter with Forex is that it isn’t a limited field problem, or at least the limits of the field are rather vast. The Forex market isn’t a linear problem, with easily definable parameters. AI has already demonstrated that it can master problems of this nature, such as chess and Go, but even the highly complex Chinese game of Go – which it was doubted that computers could master – is simplistic and limited compared to understanding markets; let alone reacting to those markets in real time with intelligence and insight.

Another problem for machines with trading is that this is a non-deterministic activity. In short, the same input can have different consequences when the overall environment differs. This is hugely problematical for computers! Machine learning relies on trial and error, but it is difficult to inform computers trading in Forex what their errors have actually been.

A computer cannot understand abstract concepts like ‘behavior’ and ‘markets’. Even a world-class chess-playing computer doesn’t even know what chess actually is! It just understands how to win! So, in that sense, it’s tremendously challenging to provide Forex-related data that machines can understand and use effectively, as they cannot comprehend the context of the data that they’ve received.

Therefore, Forex trading is tremendously tricky for machine learning systems, due to its time-dependent and non-deterministic nature. You don’t have time to sit and calculate, and you have to intrinsically understand the context of the market. This also causes problems with training and validation exercises, as applying algorithms that are produced from this process, while actually in live trading, is problematic. Live trading will be different as the selection of training sets must now be reapplied to different data.

Measuring Success

Measuring the success of any algorithm generated through training is also an issue. While machine learning trading algorithms are typically measured by their ability to generate profit, there is also a tendency to assess them on the basis of their ability to make accurate predictions. However, this can be completely useless in the real world. 99 good small decisions can quickly be wiped out by one big bad decision; this equally applies to games, as any chess player will tell you!

One possible way of circumnavigating this issue is to utilize a methodology which involves retraining the machine learning algorithm before any training decisions are made. By using a flexible window, the selection bias that is associated with an in-sample / out-of-sample set is diminished. Training then becomes a series of validation exercises, which helps to ensure that the machine learning algorithm functions properly at all times, even with wildly differing datasets. Measuring backtesting performance should also be considered important in the process of establishing an algorithm’s true worth and merit.

Such an approach is not failsafe. It is still subject to biases with curve-fitting and data-mining that tend to plague all such strategy building attempts. It is important to use vast datasets in an attempt to mitigate against this, and to conduct a large amount of data-mining bias evaluation tests, in order to eliminate the possibility of encouraging results occurring by chance.

Turning the Corner

However, some pioneers in the machine learning field are beginning to make progress with forex. After working in simulated and real-time conditions, it became clear to some experts in this research area that customizable algorithms are the key to success with Forex trading. 

When trading bots can be altered, modified, or even completely revamped, based on prevailing market conditions, the potential for success is greatly increased. Of course, this still necessitates a partnership between humans and computers, and also a significant amount of market knowledge, and potentially computer programming ability, in anyone who wishes to utilize such technology.

But the combination of human understanding and computing calculation can be absolutely deadly. When forex trading bots are aligned with occurrences in the market, they are far more able to respond in an agile fashion. It seems that this will be the way forward with machine learning and trading, at least in the shorter term.

Developing effective machine learning in Forex is far from easy, and will take a concerted effort from lots of determined individuals over a potentially long period of time. But the rewards for those who are eventually successful will potentially be massive, and will encourage a huge amount of research and development in this field in the years to come.

If you’re looking to get started on your machine learning and forex trading journey, I highly recommend these courses:

Deep Learning in Python

Supervised Learning with sci-kit learn

John DeCleene Whilst having spent a lot of his life in Asia, John DeCleene has lived and studied all over the world - including spells in Hong Kong, Mexico, The U.S. and China. He graduated with a BA in Political Science from Tulane University in 2016. Fluent in English and proficient in Mandarin and Spanish, he can communicate and connect with most of the world’s population too, and this certainly helped John as he gained work experience interning for the U.S.-Taiwan Business counsel in Washington D.C. as an investment analyst and then working alongside U.S. Senator Robert P. Casey of Pennsylvania as a legislative intern. He subsequently worked as a business analyst for a mutual fund in Singapore, where his passion for travel and aptitude for creating connections between opportunities and ideas was the perfect intersection of natural ability and experience, spending his time travelling between Cambodia, Hong Kong, and China investigating and discovering untapped investment opportunities. John is a fund manager for OCIM’s fintech fund, and currently progressing towards becoming a CFA charter holder. He loves to travel for business and pleasure, having visited 38 countries (including North Korea); he represents the new breed of global citizen for the 21st century.

Leave a Reply

Your email address will not be published. Required fields are marked *