“Diversification is the only free lunch in town”, so the saying goes… but is this still really the case? And if not, how are newer technologies changing, or more likely transforming and disrupting, some of these traditional asset management industry assumptions?

This chapter will look at some of the ongoing trends and changes in the financial services industry and how asset managers can (or should) be transforming themselves to stay relevant and “ahead of the curve”. This will include an overview of different technologies, including the buzzwords “du jour” such as machine learning and cloud computing. The predominant focus will be on asset managers active in the equity markets, as this asset class has over and over again demonstrated long-term capital growth potential; this is obviously subject to appropriate portfolio management, including de-risking and using the right technology choices at the right time.

A Very Brief History of Time and Technology

Advances in technology have been at the forefront of financial services, especially if these solutions can demonstrate a clear competitive and economic advantage. Already, back in the days of the Crimean War, competitive advantage was gained from improvements in communication speed by using carrier pigeons instead of horses. The modern-day equivalent has been to build direct microwave radio point-to-point links between stock exchanges to avoid the latency introduced by fibre-optic cables. The laws of physics suggest that this “arms race” is probably over for now. The next new kid on the block could be quantum entanglement, although this zero-latency communication concept is currently still only part of academic and philosophical society debates.

Likewise, computational processing has always been an enabling tool in most parts of financial services, and even more so within the asset management industry. Research, portfolio construction, stochastic modelling, scenario testing, probabilistic (risk) calculations and many mathematical models – including option-related calculations – are very computationally intensive.

Once again, technology is coming to the rescue: cloud computing is streamlining existing infrastructure and at the same time enabling many new, previously unimaginable or unimplementable, applications. Many FinTech start-ups would simply not exist without this democratization of computing resources. In addition to the currently available near-unlimited, on-demand cloud computing, recent progress in quantum computers could soon provide the next disruptive chapter in humanity’s unbounded appetite for computational processing.

“Quant Funds”, Machine Learning and Other Trending Technologies

Like in many other areas, it is not clear if the dog is wagging its tail or vice versa: the answer might be a bit of both with regard to technology innovation, including machine learning enabling the more ubiquitous rollout of quant funds or vice versa. Nonetheless, the trends are clearly there, with even the larger household names such as Blackrock¹and Morgan Stanley moving away from discretionary stock picking and equity analysis to more and more “software-driven” asset management decisions.

Regardless of methodologies and technologies, asset management investment objectives usually always include long-term outperformance and appropriate diversification to de-risk the portfolio and reduce shorter-term volatility. So, how can machine learning help with some of these overarching objectives?

Before delving into further asset management details, let us first explore in slightly more detail the different nuances of machine learning. Not only will this help demystify some of the concepts and terminologies, it will also help identify the key areas where more “value-add” automation can provide substantial benefits. Machine learning is sometimes more generically referred to as artificial intelligence (AI), a catch-all term that often elicits ideas of terminator-style machines going out of control. Out-of-control self-learning algorithms can certainly cause major havoc, though like everything else in life it is never black and white, however much the popular press likes to imply otherwise. Like any other new technology making its way into “everyday life”, the adoption phase and transformative implications are not always clear and certainly not in everybody’s best short-term interests. Unsurprisingly, there will be resistance from some quarters, much like the Luddites of old. Note that economic prosperity and productivity increases enabled by technology go hand in hand. This also means that companies which are slow to adopt such technologies may find themselves at a competitive, and therefore a financial, disadvantage.

Supervised and Unsupervised Machine Learning

Probably the more ubiquitous machine learning technology out there, and the one most associated with “self-learning” and “self-reinforcing”, is known as supervised machine learning. These algorithms rely on some human element to label the input data (i.e. add some structure to the data) and the machine learning algorithms can then subsequently be trained, with ideally as much relevant input data as possible. Typically, the input data is segmented into learning data sets and validation data sets. Once properly trained, the live out-of-sample performance of these trained algorithms will ideally provide some robust long-term predictive capabilities subject to some (ongoing) tweaks and retraining. Very good examples of supervised machine technology applied to real-world problems include “speech character and image recognition” and “search and recommendations”, especially in e-commerce.

The predictive powers of these well-trained algorithms will depend on numerous factors, though generally speaking performance will improve with more quality input data, as long as there are no “regime switches”. In the above examples, somebody starting to speak in a new language when the algorithms were only trained on the English language would be a “regime switch” most probably requiring new training data. Another limitation of supervised technology is the inherent assumption that future patterns are an amalgamation or a repetition of historic data.

Unsupervised Machine Learning

The unfortunate reality is that a lot of financial time series are very prone to regime switching and the past is not always a good (or indeed any) indication of future events. Examples of regime switches could include changes in interest rates (even negative interest rates these days!), never seen before quantitative easing programmes, new regulations and the introduction of new financial products (CDOs, etc.). Generally speaking, the more random and less repetitive the input data, the more difficult it is to learn anything from the past.

Thankfully, not all machine learning algorithms require historical data set training, human labelling and non-regime switching data universes. In addition to supervised machine learning technology, its younger, more recent unsupervised machine cousin is exceptionally good with the realities of financial markets. Working on the assumptions that Brownian (“random”) motion models are applicable to describe some of the pricing behaviour of traded financial instruments, can we use unsupervised machine learning for forecasting capabilities? The short answer is a big “yes”, although only under some circumstances, specifically when financial markets are experiencing “regime” switches and moving away from non-Brownian motion states. Using limit-order book information from global financial exchanges, complex clustering and characterization algorithms and distributed cloud computing, AlgoDynamix is providing forecasting solutions to asset managers and other global financial institutions all around the world.

Putting it All Together, Portfolio Sciences for the 21st Century

So, are Smith’s invisible hand, the efficient market hypothesis²and portfolio diversification still applicable concepts moving forward? Probably yes, but like everything else, a lot of things work a lot of the time until they don’t… Extreme events in financial markets, aka tail risk losses (or returns!), have historically been impossible to predict with any sense of consistency, and most of the money is either made or lost during these extreme events.³ Moving forward, it is now possible (and even more important) to stop using predictive technologies based on existing models and past events. Newer technologies, including the AlgoDynamix directional market risk forecasting solutions, are already being used by asset managers and other financial institutions globally. Looking at the current FinTech revolution and the increasing rate of technology adoption in all areas of financial services, the future does look good for the right types of AI in the right places.