Yves Chauvin of Axa Investment Managers' Rosenberg Equities considers the rate of technological progress in areas such as machine learning and ‘big data’ in asset management.
Machine learning and big data have seen a surge of interest in recent years, as a result of various drivers. On the technology side, there have been significant strides in computing power and data access that have pushed the boundaries of what is possible with these techniques. From a publicity perspective, several high-profile events have captured the public’s imagination, helping to broaden the interest in the topic. This includes Google’s development of a self-driving car and the victory of its artificial intelligence programme against the world champion ‘Go’ player.
The asset management industry has also seen significant technological progress and the increased availability of data. However, I believe it’s fair to say that it is still somewhat behind other areas in terms of what has been achieved. The success of artificial intelligence in highly structured environments like Go, for example, is perhaps not entirely surprising given the progress in data processing, but we’re yet to see whether machines can learn some of the higher-level functions of investment decision-making in highly complex markets.
So far we’ve seen analysts try to forecast dividend cuts using what are known as ‘random forecasts’. These are collections of decision trees that seek to estimate the likelihood of a cut for a particular company by splitting the ultimate outcome – to cut or not to cut – into a subset of related, binary, outcomes and weighing the probability of each.
Another, earlier example was the use of neural networks – a system inspired by the way the human brain gathers and processes new information – to identify future merger and acquisition targets. If you can predict takeover targets, you should also be able to profit from that knowledge. However, the original approaches were susceptible to spurious patterns in the data and now serve as the basis for more refined techniques.
Learning techniques utilising neural networks and other frameworks can be seen as extensions of non-linear regression and parameter estimation formalisms. There is a lot to be gained from seeing these techniques in the classic light of maximum likelihood, Bayesian calculus, signal detection theory and statistical theory. Mathematical formality is guiding us towards proper model architecture and appropriate innovation.
HOW TO AVOID SIMPLY ‘DATA MINING’
As the saying goes, if you torture the data long enough, it will confess anything. This is a particular challenge in a field like investing because it is not a ‘natural’ process. Market prices are not set by nature, so the data is typically ‘noisy’ and there is a risk of the machine learning the noise as opposed to the signal. The risk is particularly acute where the human role in the process is reduced, such as in Deep Learning. Deep Learning is the branch of machine learning that minimises human pre-training of models, essentially building a framework to guide the computer in its learning process. It emphasises computers’ ability to teach themselves by discerning patterns in large amounts of data.
Experience and knowledge of a domain and that of the underlying data is the ultimate protection against simplistic data mining. Like any other scientific endeavour, it is both the successes and failures that should shape the investment process to have a better understanding of equity valuation, and can therefore guide machine learning investigations.
At the moment, machines need humans. Perhaps in the future they won’t, but they do now. In general, I think there are limits to the application of Deep Learning. I doubt that computers can learn everything simply by being let loose on giant data sets. The analogy I draw is that machine learning is a little like the learning children do before they go to school – it’s impressive, but limited.
School is where you really develop and deepen your knowledge because you are being taught by people who know.
Currently, machines ‘learn’ directly from human experts. The exploration of direct machine learning of ontologies – using natural language processing, a field focused on the interaction between computers and human (natural) languages – has just started. The integration of machine-learning techniques makes it exciting to have both humans and machines working together.
ALFA IS IN THE OUTLIERS
Investment managers are increasingly using quantitative techniques and analysis when addressing the needs of investors. They should therefore also see themselves as knowledge managers to an extent. The aim is to constantly increase the knowledge investment managers hold and to integrate that knowledge into software that allows them to make the best use of it.
In practice, that means translating the fundamental insights and intuitions of researchers, portfolio managers and traders into computer code and rigorously testing those insights and intuitions. Data is a central, but in my view not sufficient, component in that process. And the quality of the data is as important as the quantity because the alpha, like the majority of errors, is in the outliers.
We have been integrating and refining financial data for over 30 years, in part to make sure outliers are not due to simple data entry errors. Understanding the story behind data outliers within a context leads to investment insights. These insights in turn have the potential to show which machine learning techniques are appropriate. I do not believe data or machine learning techniques will become commodities in the investment domain unto themselves. The value of the investment process still resides in a deep dive into the data and in the complete integration of the investment process from data collection through to machine-learning processes.
So, while increased data availability and machine learning are exciting, the analysis completed in areas like non-linear regression would not have been possible a decade ago and neither is it likely to be a silver bullet. There is still a lot of work to be done to determine how all of these techniques can be best harnessed to achieve the objectives of investors, but we are certainly witnessing the ‘rise of the machines’ in asset management.
Yves Chauvin is director of the investment data platform at Axa IM Rosenberg Equities
©2016 funds europe