Artificially Intelligent Financial Trading

The Rise of the Artificially Intelligent Hedge Fund, on Wired magazine's website, describes a new hedge fund running solely on AI decision making. There is no human discretion at all that can modify bad decisions, which are all made by machine learning techniques. At least, that is what the article implies but I know that all algo-trading systems have a circuit breaker for those inevitable Black Swan days.

For me the article is particularly interesting because it mentions evolutionary computation, which I describe in my recent article EDDIE Beats The Bookies, detailing my own research. The company building this AI financial trading engine is called Aidyia and uses both evolutionary computation and probabilistic logic.

The fund made 2% on its first day of trading. Most sports algo-traders would be happy with that figure. Even the majority of manual sports traders would be happy with that figure too. Other companies using AI for fund management include; Sentient Technologies, Two Sigma and Renaissance Technologies.

Many trading firms use teams of quants to build computer models to decide where to place funds but there is always a degree of human discretion involved. Removing human discretion is a natural progress as it removes emotion, human error and speeds up the trading process to get the best prices available. Other automated trading systems will be competing in the market and the first to act is the first (often the only one) to profit.

Computer (quant) models tend to be rather static and do not adapt to changes in market sentiment too well due to the human input required to update them. AI systems are adaptive and run in real time, constantly updating as markets change and evolve. Indeed, AI systems also try to predict these changes before they occur, which allow for impressive market coups.

By using evolutionary computing, a population of virtual stock traders are evaluated, the superior traders pass their genes on to a new generation of traders whilst the old one is culled. This process continues until the population of traders learn traits that adapt them to become profitable traders. Whenever these traders start losing money then it is assumed the market has changed and new virtual traders can be evolved to replace them.

In sports trading I use a similar process and create a population of sports traders that evolve traits that allow them to generalise about movements on Betfair and BetDAQs exchanges. The task is never-ending with hardware and software changes to decrease latency, increase speed and updates for the latest technology; price and order streaming being the latest.

Further Reading

Programming for Betfair is a guide to creating applications for direct access to Betfair's exchange and will therefore be useful to those wishing to implement an algorithmic trading set up using the other books listed here.

No previous programming experience is necessary to build the applications in the book. After completing the programming exercises the reader will have a powerful tool for gathering prices for database creation, strategy building and algorithmic trade placement. Beginner programmers and experienced programmers have informed me that the book is easy to understand and that it has assisted them in creating algorithmic trading platforms.


  1. Hi James,

    I've been doing a fair bit of research into evolutionary algos and their implementation in R. What I'm struggling to understand is whether these algos can be used in themselves to predict a continuous variable (e.g. a regression task) or whether they are used as a feature selection you mind elaborating on how one would implement them in a regression setting?

    1. Hello Antony,

      A search on Google Scholar reveals plenty of papers on GA Regression Analysis so it is being done. It is not something I do but I can imagine it would be useful for fundamental analysis of sports data.

      Personally, I use Genetic Programming for classification problems.

      If you want to know more about GA applied to regression then I suggest that you read academic papers and look for online tutorials.

    2. Thanks James, we are still 3-4 years behind the rest of the world here in Australia in terms of working with sports data. Linear regression is still the tool of choice for most so there is an edge when using newer AI algos such as deep learning and GA.

    3. I've looked at this "deep learning" but all I see are the same old neural networks.

      I prefer genetic computation because you can see the "thinking" behind the solution.

  2. Interesting, the deep learning models I've been using as of late have been significantly out-performing the old classical statistical models I had in place. Will be interesting to see if GA can set a new benchmark for me.

  3. Actually a possible merger of the two approaches which may yield good results is the use of GA to transform the weights or parameters of the deep learning algo....

  4. Hi James, good to hear that Eddie is making a comeback, its a paper I have fond memories of. From an automated trading perspective I just wanted to check whether the stability of Betfair and the API has ever been an issue. Perhaps you have some suggestions on this.

    1. Hello Mark,

      The new API-NG is very easy to program to compared with the previous API-6.0

      There were a few issues when API-NG came out beta but they were all quickly dealt with and everything now runs better than the previous API.

      For the self-coder more calls and more data can be downloaded than before, making data capture and multi-market trading all the more easier.

      Not to mention that you can build an application tailored to your needs without giving your secrets to third-party programmers and you don't have to pay any subscription fees.