The significant transformative impact that artificial intelligence (AI) will inevitably have on the world is continually becoming more widely acknowledged.

With this, investment management companies of all ethos and sizes across the globe appear to be in a race to reap the promised benefits of alpha and efficiency being offered by the application of machine learning technology.

Stuart Reid, chief scientist at machine learning powered unit trust NMRQL Research, says that while this keen sense of competition will certainly help in driving forward the machine learning capabilities of the industry as a whole, it is important for prospective investors to be aware that not all machine learning models are made equal, particularly when it comes to navigating the high volatility of financial markets.

“Most off-the-shelf machine learning models assume that the data generating process is stationary, which is to assume that the behaviour being analysed and ‘learnt’ doesn’t change. While this method may largely work for stationary systems like mapping roads and image recognition, the reality is that financial markets are inherently nonstationary and require a machine learning model that is able to continuously evolve with whatever volatility they may experience.”

This highlights the importance of building a system that continuously monitors how well a particular model’s beliefs and behaviours map to the real world, and adapts to change, explains Reid. “As the model will never be sure of what is going to happen, given the high level of noise in financial markets, uncertainty needs to be taken very seriously. This is done through predictions and distributions of confidence based on probability functions.”

Based on this, Reid points out a few red flags to look out for when choosing which machine learning management fund to invest with.

* Stationary models – “Firstly, if a company is using an off-the-shelf solution that is based on historical data and lacks the ability to evolve and adapt in real-time to the changing markets, that’s a massive red flag,” explains Reid.

* No uncertainty measure – “No prediction in financial markets can ever be 100% certain, so a solution needs to be able to measure the certainty of the model and provide feedback when change occurs or model error strikes.”

* Quantitative investing, rather than machine learning investing – “The difference between quantitative investing and machine learning investing goes back to the ability to introduce a fundamental feedback loop between the market and the model, thus creating a mechanism where the model will constantly adapt at the sign of bad performance.

“A quantitative investor will therefore build a static model of the world as they see it, which will inherently incorporate some personal biases and beliefs, whereas a machine learning system looks at data and builds its own dynamic and responsive view of what the world looks like, without any preconceived assumptions or personal biases.”

Reid concludes by providing some valuable insight into the current machine learning methodology being applied at NMRQL. “We developed a general-purpose, multivariate time series prediction platform which allows us to easily create thousands of collaborative supervised online learning agents which collectively encode self-organising investment strategies which adapt alongside the market. What this essentially means, is that we developed a dynamic learning machine model which allows the investment strategies we employ to constantly adapt and self-correct alongside current market movements.

“As of March 2018, we have about 1200 deep learning algorithms deployed in production which collectively process 19000 independent time series and produce hundreds of GB of information a week. This will continue to grow because innovation is the bedrock of our investment philosophy.”