Machine Learning (“ML”) has been a relatively niche field for decades, with interest stemming mainly from academia and the life sciences. In recent years however, the development of ML technology has started to disrupt and, in some cases, reshape industries. This is evident in the hedge fund sector, which has seen a plethora of ML funds launching, or expected to launch, in the first half of 2018.
ML can generally be defined as a subset of artificial intelligence. It involves using statistical techniques on large amounts of data to learn how to perform a specific task or find patterns in order to make predictions and solve problems. These models optimise automatically through experience and with limited or no human intervention. In reality however, the term “machine learning” is as broad as the fields of mathematics or statistics.
While we may not be aware of it, ML tools are being used in a large number of everyday applications. This includes social networking platforms such as Facebook’s facial recognition technology, recommender systems used by Amazon and Netflix and in virtual personal assistants, such as Apple’s Siri and Amazon’s Alexa. However, they are also having a considerable effect on established industries including healthcare, education and finance.
Within the hedge fund industry, ML techniques are used in two different ways to increase alpha: through signal generation and as part of the investment execution process.
The majority of funds which use ML techniques for signal generation are either statistical arbitrage or long/short equity strategies, with many adopting a market-neutral approach and a focus on US financial markets, such as the S&P 500 equity index.
Deep learning is a common form of ML that is used within these strategies. It uses algorithms that work in “layers” – for example, by analysing relationship patterns between financial statement data, governance changes and press releases to a security’s historical price. Another example of this technology is in satellite imagery, where ML techniques are used to monitor shipping routes or store car parks, in order to create growth and consumer demand predictions. Other techniques include genetic algorithm modules that employ a survival of the fittest approach. Parent models are combined in different ways with the intention to create superior child models with a higher predictive accuracy.
In contrast, funds which use ML techniques for execution tend to be traditional quantitative funds that are gradually building out advanced execution technology. This technology will create short-term price predictions, based on historical tick data trends, and has generally improved net performance for managers employing these techniques, in some cases by up to 2% per annum.
Expectations surrounding the potential of Machine Learning strategies are high, with some calling it the “third wave” of investment management.
Expectations surrounding the potential of ML strategies are high with some calling it the “third wave” of investment management. The main differentiator of these strategies is the ability to define their own rules and update independently to changing market environments, in comparison to traditional quantitative funds that use static, predefined models. A further differentiating factor lies in the ability to construct relationships in multiple dimensions and extract patterns which were previously not visible to the human eye, potentially generating alternative sources of alpha. These strategies also benefit from their vast data processing and analytical capabilities on structured data, including price and fundamental data and unstructured data, such as job posts, social media feeds and data obtained from mobile devices. Many of these ML funds are targeting an attractive low to mid double digit return, however owing to the fact many of these strategies only recently launched we are yet to see whether this return profile will play out in reality.
While these strategies may sound attractive, there are challenges to their success. Firstly, the majority of ML experts do not work on Wall Street, with a vast number employed in education, high-tech and telecommunications industries. In addition, when ML technologists with limited experience in financial markets create models based purely on historical simulations, they are prone to overfitting, creating exceptionally strong backtests, while these relationships may not hold up in a live trading environment. ML techniques are also required to train on vast volumes of data in order to make future predictions. Therefore, issues may arise due to the relative recency and limited quantity of financial data sources, from which there is a limited amount of data with which to train the machine. Due to the sophistication of these techniques there is also the possibility of these models taking on a “black box” nature, however hedge fund managers are generally aware of this issue and hence mainly follow a hybrid investment approach with an element of human intervention.
There are evidently barriers to success in ML. When it comes to using ML for investing capital, driving a car or using autonomous robot surgeons, the stakes are much higher than using ML in safer areas such as facial recognition or recommender systems, and hence the hurdles to mass adoption are also much higher. Therefore, unsurprisingly ML in these areas is further behind in terms of its real world application.
The use of ML strategies as part of the overall investment process offers substantial promise to the financial industry if the specific risks are properly managed. We expect hybrid approaches, combining both man and machine, to remain popular until ML strategies demonstrate their ability to generate alpha and stability through market cycles. However, with a rapidly expanding universe, ML funds are here to stay. With this in mind, investors will be keen to gain an understanding of the idiosyncrasies of ML strategies, and an awareness of the advancement and success of this technology in practice.