Views expressed are those of the author and are subject to change. Other teams may hold different views and make different investment decisions. The value of your investment may become worth more or less than at the time of original investment. Any forward-looking estimates or statements are subject to change and actual results may vary. While any third-party data used is considered reliable, its accuracy is not guaranteed. For professional or institutional investors only.
Man has long dreamed of creating artificially intelligent (AI) machines, both for the intellectual challenge and for their potentially profound usefulness to society and industry. However, success has proven elusive — until recently, that is.
Beginning circa 2012, progress accelerated sharply, yielding impressive AI successes in domains such as image recognition, speech recognition, language translation, game playing, and autonomous driving. In the space of just a few years, the capabilities of computers in these and other domains have advanced from laughably poor to superior to those of an average human.
The latest milestones
A recent and powerful example of this progress was the late 2018 announcement by Google’s DeepMind of “AlphaFold,” an AI system for predicting the shape of proteins from their genetic composition. This “protein folding” challenge has been of long-standing interest due to the importance of this information in understanding biological processes and designing new drugs. Google’s system placed first in a 2018 biennial competition, with its superiority over prior methods termed “unprecedented” by the event organizer.
In addition, in 2020, a team of MIT researchers leveraging a deep-learning algorithm identified a promising new drug candidate for treating bacterial infections. Dubbed “halicin” after the artificially intelligent “Hal” of “2001: A Space Odyssey,” the predictions of the machine learning model were validated by tests in mice which showed high effectiveness, even against antibiotic resistant strains.1
These are potentially significant milestones, for not only their narrower areas of study, but also more broadly as further proof that the rate of progress in AI/machine learning is more likely accelerating than plateauing. This acceleration could well have major implications across many industries and society as a whole. To some extent, it already has.
Applications widespread across industries
Stories are now appearing on an almost daily basis describing new applications of AI across a broad range of critical tasks. For example:
- Drug design (e.g., applying protein folding)
- Computer vision (e.g., face recognition for social media)
- Autonomous or computer-assisted transportation (e.g., self-driving cars)
- Interpretation of medical images (e.g., detecting cancer)
- Applied pattern detecting (e.g., fraud in credit card transactions)
- Language translation/text summarization
- Computer security (e.g., detecting malware, filtering spam)
- Speech recognition (e.g., in customer service applications)
- Personal assistant (e.g., Apple’s Siri or Amazon’s Alexa)
- Algorithmic trading in financial markets
- Computer gaming
- Industrial optimization (e.g., using machine learning to lower data center cooling costs)
- Marketing (e.g., forecasting, market analysis)
- Creative arts (e.g., photo manipulation, artificial art)
I believe these emerging AI capabilities will be important and, in some cases, transformative across industries. In the case of Google, its success and actual existence is largely due to superior AI (from its page-search algorithms), and it has publicly proclaimed itself an “AI-first” company. For many other companies, AI will help enable the delivery of superior products and services at lower cost. Ultimate competitive success will likely be significantly impacted by the alacrity with which firms embrace and invest in AI.
Investment management relevance more nuanced
Although I strongly believe AI will be of great relevance to most industries, its applicability to investment management is more nuanced. The particular challenge of security selection is very different and more difficult than the other tasks listed above, for several reasons:
- Relatively little data is available in finance for machine learning. Although we have thousands of securities observed over hundreds or thousands of months, they tend to commove significantly at times and thus are not useful independent data points. Also, major economic events tend to play a big role in driving security movements. Therefore, machine learning techniques are prone to “overfit” the data set, finding weak patterns that mostly represent random variability rather than true investment opportunity.
- The available data is extremely “noisy,” with a lot of random variability contaminating potential true patterns in the data. This contrasts with domains such as computer vision, where an image is almost all signal and no noise.
- The economic and financial worlds are constantly changing due to evolving societies, industries, governments, etc. This makes learning from past data potentially perilous. In contrast, tasks like robotics are grounded in immutable truths, such as the laws of motion.
- “Trial and error” learning is not practical in investing (unlike, for example, the task of a robot learning to walk). We act as fiduciaries and cannot “experiment” with our clients’ assets. Furthermore, the time frame over which we receive feedback and can adjust our actions is typically far too long (weeks or months, at least) to be practical.
- Investing is an adversarial domain. Therefore, it is not safe to take actions based purely on past data without also imagining what the future competitive response may be.
Bottom line: Using advanced AI methods for stand-alone, black-box security selection is probably not their most suitable application. Having said that, the knowledge and use of modern AI may present significant opportunities for investment managers, but in more subtle ways:
- Systematic investors can continue to apply the statistical insights provided by machine learning in seeking to identify and exploit inefficiencies in asset pricing. However, the applications will be guided by scientific judgment and domain knowledge about how markets are likely to function.
- Fundamental investors will collaborate with their data science colleagues to collect and harness the incremental insights and data sources available from the new generation of big data and AI techniques.
- Fundamental investors will also leverage AI to help form their assessments of industry dynamics and the companies they cover, potentially enabling them to better identify the likely “winners and losers.”
The impact of technology in today’s world is easy to see. What is harder to predict is the impact that technology could have in the years and decades to come. The evolutionary path that machine learning and AI follow — or more likely, lead — could have lasting ramifications for labor markets and economies globally, as well as for specific industries and corporations. (Forward-looking investors, take note.)
1“Artificial intelligence yields new antibiotic,” MIT News, February 2020.