Zafrin Nurmohamed, Butterfield

While 2018 was a difficult year for investors, following a tumultuous fourth-quarter drop of 13.5 percent in the S&P 500 and a full-year decline of 4.4 percent in the S&P 500 Total Return Index, quantitative hedge funds managed by the likes of Bridgewater, D.E. Shaw, and Renaissance Technologies, performed comparably well.  In 2018, Bridgewater’s flagship Pure Alpha Fund gained 14.6 percent, D.E. Shaw’s Composite Fund gained 11.2 percent, and Renaissance Institutional Diversified Global Equities Fund gained 10.3 percent.

Underlying the performance of quantitative hedge funds are algorithmic trading strategies that sift through data searching for exploitable market dislocations. A number of hedge fund managers are experimenting with machine learning techniques such as sentiment analysis, an application of natural language processing, to determine whether, for example, a higher proportion of negative words in public statements made by central bankers and CEOs portend worsening economic data and/or share price performance.

Others leverage satellite imagery to determine how much oil is being shipped or how busy parking lots are outside of specific retailers with the goal of predicting sales, and ultimately earnings.

The investment industry, like others, is increasingly being disrupted by advances in machine learning.  The analogue of self-driving cars in the automotive industry may very well be self-driving portfolios in the investment industry.  Indeed, robo-advisors, which leverage computer algorithms to automatically make asset allocation decisions based on historical data, have increasingly attracted investments.  While self-driving cars may ultimately prove to be safer than human drivers – at least most of the time, when the surroundings become different from what’s been observed in the past, autonomous vehicles may find themselves getting into accidents.  Having a safe pair of hands at the wheel to take over could mean the difference between a bad accident and safe passage home.

In their book “Prediction Machines,” Ajay Agrawal, Avi Goldfarb, and Joshua Gans argue that in a world where the cost of prediction becomes cheap due to advances in machine learning, complements of low-cost predictions, namely judgement and data, will become increasingly valuable.  In 2012, when European Central Bank President Mario Draghi commented that the ECB will do whatever it takes to preserve the euro, seasoned investors quickly took notice and used their judgement to start buying up government bonds causing nominal yields to fall, ultimately to negative levels in several European countries. Such levels were not justified by economic relationships generally taught in prevailing textbooks – imagine, paying someone to take your money – and many algorithms sifting through past data, especially those which overlooked Switzerland’s experience with negative interest rates in the 1970s (or didn’t have access to that data), incorrectly predicted that nominal interest rates would rise as they fell to zero when in reality a new regime had materialised.

So how can investors taking advantage of machine learning models protect themselves from losing money, especially when the surroundings look different from the past?

Examining assumptions, or priors, and training data underlying these models are good first steps.  Some machine learning models, however, are easier to examine than others. Linear regression models, for example, require humans to choose which variables, or features, to include in their models and the predictive power of these variables can be monitored over time. If the variables chosen no longer matter, e.g., because of a regime change, humans can choose to turn the models off.   In other machine learning techniques, such as deep neural networks, which are often compared to black boxes, it’s more difficult to trace back predictions to one or more input features.  In such cases, it’s harder to understand why these models are predicting what they are. And, more difficult to determine is whether these models have become irrelevant in a new regime and whether they should be turned off (current research efforts are seeking to make deep learning models more interpretable).

Absent of regime changes and new surroundings, however, machine learning models are powerful tools that can be taken advantage of. When the surroundings look different, a safe pair of hands coupled with machine learning algorithms can prove to be an unbeatable duo.  Like the advent of the spreadsheet, machine learning advances will enable investors to make predictions under many more scenarios than ever before.

Looking forward, many economic unknowns confront both the real economy and the investment world including the impact of Brexit, trade disputes between the US and China, and how central bank balance sheet tightening may affect asset prices.  While major central banks have recently become dovish, including the Federal Reserve which recently indicated that it will explore ending its balance sheet runoff, it is unclear whether persistently large balance sheets will support asset prices in the same manner as they have over the past decade.  For quantitative hedge funds, and other tech savvy investors alike, a key set of investment decisions for 2019 will include whether to leave machine learning models running on auto-pilot.

Sources:  CNBC, Financial Times, Impact Lab, Knowledge @ Wharton, MIT
Disclaimer: The views expressed are the opinions of the writer and whilst believed reliable may differ from the views of Butterfield Bank (Cayman) Limited.  The Bank accepts no liability for errors or actions taken on the basis of this information.