Artificial intelligence has recently experienced increasing attention, following staggering achievements in applications such as self-driving cars. This article shows how the adaptivity and self-learning capability of machine learning tools may add value for asset managers. However, the inherently flexible nature of machine learning methods is also their biggest challenge. It requires that the methods are put in the right context and thoughtfully applied. We are convinced that machine learning will most likely turn out not to be the much-searched Holy Grail, but that it will help quantitative investment managers in further improving their allocation processes.
From machine learning in general…
Machine learning refers to extracting knowledge from data by identifying correlated relationships without getting prior information about what causal dependencies to look for. It was as early as 1956 that John McCarthy coined the definition of artificial intelligence as “the science and engineering of making intelligent machines”. However, it has been mostly due to significant advancements in computing power and data availability that the application in recent years has become applicable for everyday life.
The currently most successful field is supervised learning, where algorithms learn based on provided training data that reveal known relationships. Examples include tasks like the detection of fraud in credit card transactions or the creditworthiness of debtors. Unsupervised learning algorithms on the other hand only receive input data to learn from, but no information about the output data or relationships. They detect patterns in the data by identifying clusters of observations that depend on similar characteristics. Machine learning can for example be used to identify the main topics in the news flow for a given stock. Combining methods of supervised and unsupervised learning results in the so-called reinforcement learning, where the algorithm first detects patterns on its own, and then receives feedback from an exogenous source to validate or further guide the learning process.
Deep learning or neural network methods mimic the function of the human brain by feeding information through different layers and nodes. The simplest form can be seen as a generalization of linear models that perform multiple regression steps.
… to specific applications in quantitative asset management…
n order to show models that are appropriate for specific tasks, the investment process can be subdivided into various steps. For example, the definition of the investment universe. Here, machine learning tools may add benefit by identifying uncorrelated assets that provide true diversification benefits, or by mapping data into new representations that allow for other interpretations such as the detection of style drifts in hedge fund strategies or factor exposures such as momentum or value.
Dendrograms, for example, can be used to structure a set of individual commodity markets into more meaningful clusters. Ideally, it comes up with the well-known sectors such as energy, precious and industrial methods.
At the bottom of the visual representation in figure 3 are the single data points that are joined in first clusters. The model groups copper and aluminum to a mutual cluster of industrial metals, or gold and platinum to precious metals. These two clusters are then joined to form a cluster of metals in more general. Similarly, heating oil and crude oil are merged before being clustered together with gas oil as the energy complex. The energy and metals cluster are then put together to form a cluster of commodities that are more dependent on business cycle swings. The soft commodities soybeans, soybean meal, corn and wheat are structured in a separate node that only consists of agricultural products. Interestingly, the natural gas commodity forms an individual cluster, most likely because of seasonality factors that separate it from the other energy commodities.
… and challenges and limitations
Artificial intelligence works especially well for tasks with precisely defined rules and stable probability distributions. However, the stochastic nature of financial markets with its lack of stable rules and probability distributions may challenge the validity of relationships that are learned from the past. Accordingly, models should always be applied to clearly defined problems and validated against sound theoretical assumptions Additionally, despite the seemingly abundant access to data, there is only one historical price trajectory for each financial market to train a model on. This limited data availability restricts the complexity of the artificial intelligence model that can be applied and therefore the flexibility of its output when forecasting future price movements. As a consequence, researchers should focus on parsimonious model structures and not be misled by the mightiness of artificial intelligence models to adaptively learn the past.
Urs Schubiger, Egon Ruetsche (above left and right) and Fabian Dori (above) are quantitative strategists in the Systematic Trading Group at Aquila Capital