May Newsletter: The AI Revolution

Posted by Alex Frey (@alexhfrey )

In 1997 the computer "Deep Blue" beat Chess Grandmaster Garry Kasparov in a best-of-seven match, proving machine supremacy over the best human chess player for the first time in history. Deep Blue was the culmination of more than a decade of work from a team that included its own chess grandmaster. Earlier this year, Google came up with an algorithm that within 24 hours of 'training' exceeded the level of all humans ever at playing not just Chess, but also "GO", an even more complicated game that requires more than just "brute force" calculation.

These advances have been made possible by faster computers and larger datasets which have allowed researchers to train larger and larger "neural nets" -- a type of machine learning architecture that enables computers to learn to "put the pieces together" to accomplish tasks that require many layers of abstraction. For instance, early in a neural network trained for image recognition, a particular computational unit could serve to detect edges in a picture -- later on a neural network could work off the edges detected in these earlier computational units in order to find which of those edges when put together represent a human face. This may be similar to how a human brain works -- when given an image we do not analyze every pixel individually but instead recognize lower-level patterns like edges and medium-level patterns like faces in the process of building our mental picture of what an image represents.

The fundamental breakthrough is that when there is a large enough dataset to "train" an algorithm with, the new methods no longer rely on a human programming logic into the computer to tell it what to look for. Instead, a neural network architecture is setup and a computer literally learns what it needs to learn in order to optimize performance at a given task. Rather than telling the computer what to do, the human instead gives it an objective, and sets some architectural parameters (how many layers / nodes as well as choice of several tuning parameters) and lets the machine learn on its own. The result, as the "chess" and "go" examples show, is that the computer can come up with strategies on its own that are initially entirely mysterious and counterintuitive to even the best human players, therefore besting them at their craft.

A very concrete example of this is language translation programs, for instance a program like google translate that takes a sentence or paragraph in English and translates it into any number of other languages. One approach to this type of translation is to program the rules of grammar into a computer along with dictionaries. This relies on humans to supply the machine with much of the logic. The deep learning approach is to simply give the machine a "training set" consisting of a series of translated paragraphs (these could be from books that have been translated into multiple other languages by humans, for example), train a neural net, and let it figure out all the logic for itself.

This has of course all led to considerable speculation, hype, and, unsurprisingly, fear that the machines will soon take over. There is no question that advances in AI are going to have a big effect on many aspects of the economy. The self-driving car, when it arrives, will bring enormous societal changes in itself. More broadly, the AI revolution could lead to more rapid growth in labor productivity, and rising productivity is the ultimate source of economic prosperity. Also, it could help the developed countries of the world overcome the demographic problems associated with an aging workforce. So artificial intelligence could be a boon to the economy, to stocks, and to investors. Of course, there is always tension between rising productivity and unemployment, but the economy has successfully dealt with this problem since the beginning of the Industrial Revolution, and it may well do so again.

In this newsletter, we want to address a much narrower question: if AI is already better at chess and Go then even the best human players, is there any reason it soon won't be better at investing too? Will human fund managers go the way of the dinosaur, replaced by AI algorithms that learn to pick investments on their own? And what does that mean to us as individual investors?

To begin with we need to distinguish between algorithmic investing, machine learning, and "deep learning" (which is the field that uses the neural nets described above). Algorithmic investing uses algorithms designed by humans to allocate funds among asset classes or to pick specific investments based on available data. In this case, a computer is following human designed rules. The computer speeds the process and eliminates some human biases, but it is basically just automating what a human may have done manually in the past. Most robo advisors and smart beta funds are practicing algorithmic investing. Of course, IvyVest is also performing a form of algorithmic investing. Our algorithm is certainly systematic and "rules-based", but the rules were designed by us, not uncovered by an AI set free to uncover the best strategies on its own (there are reasons we will get to that explain why this would not work particularly well when it comes to asset allocation). The simplest form of algorithmic investing is perhaps the index fund. Index funds have an algorithm that tells them to simply buy a capitalization-weighted version of the benchmark. The terminology here can get confusing because in a way the index fund is the furthest thing imaginable from what most people think of when they think of "artificial intelligence", but because it is still a form of systematic investing index-like strategies can sometimes get lumped in the generic "quantitative" bucket.

Machine learning goes a step deeper than algorithmic investing. The computer is programmed to ingest large amounts of data and alter its algorithms based on trends that a human might have difficulty finding. The ability of the computer to quickly ingest and analyze large amounts of data gives it a significant advantage over humans, but human programmers are still more in control of designing the particular "features" that a machine will look for (for instance in designing a stock-picking app you might suggest that the price to free-cash-flow ratio of a stock is one of many indicators that, when put together, express whether the stock is a good buy or not). In this case, humans can still in many cases understand the resulting algorithms.

Deep learning puts the program in charge of itself. Based on so called neural networks, the computer ingests large amounts of data and modifies the controlling program in ways that a human might not be able to understand. Deep learning has been very successful with many complicated tasks, such as language translation. It seems likely to be helpful in evaluating the great variety of information that goes into choosing stocks or quickly responding to market conditions.

We would like to be able to tell you how AI is currently being used in investment management and how successful it has been. Unfortunately, it is difficult to get information on exactly how managers are currently using AI. There are many “quant” funds that pick their investments based on quantitative algorithms, but it isn’t clear how many of them are truly using machine learning or deep learning, and because of the hype surrounding the field, it's more likely than not that anything they are doing will get branded as "AI" if it is helpful for marketing purposes.

However, "quant" funds (which are based off machine learning or deep learning) seem to be gaining assets at the expense of more traditional fund managers. Bloomberg news recently reported that non-quant hedge funds had net outflows in 2016 and 2017 while quant funds had net inflows. At least one quant fund, the Medallion fund run by Renaissance Technologies, has been astoundingly successful. Figure 1, taken from reference 1, shows how the Medallion 2 fund has performed compared to the S&P 500. Note that it had a large increase in performance in 2008 while the S&P 500 was collapsing. Unfortunately, we don’t know whether Renaissance is using true AI (machine learning or deep learning) or whether they are using more traditional human designed algorithms. The Medallion Fund is available to a very restricted number of investors, most of whom are Renaissance employees.

img1.png Figure 1: The return of the Medallion 2 fund compared to the S&P 500. Data was taken from reference 1.

There is one fund that advertises that it is using AI and is available to average investors. The “AI Powered Equity ETF”, stock symbol AIEQ, claims that it picks stocks using the IBM’s Watson. It is a new fund that started in the fall of 2017. Figure 2 shows its performance to date, but of course it is much to early to tell if this fund will be successful. In the early going, it is beating the S&P 500 by a small amount.

img2.png Figure 2: The performance of AIEQ (an AI powered fund) compared to the performance of the Vanguard 500 Index Fund (VFIAX)

How are the other quant funds doing? Eurekahedge, an independent data provider, maintains hedgefund indices that track various types of hedge funds. One of their indices tracks funds that use AI in their stock selection. Figure 3 shows how the performance of the AI index compares to Eurekahedges’s general hedge fund index and to the S&P 500 (represented by VFIAX). The AI index is very slightly behind VFIAX over this period.

img3.png Figure 3: Comparison of an AI Hedge Fund Index with a general hedge fund index and the S&P 500 (represented by VFIAX).

If computers are already so much better at humans at complicated and brainy games like Chess and Go, why do they, at least according to these results, not seem to have reached their potential when it comes to the markets?

One factor is likely that true deep learning / AI requires a really large amount of data. Popular image recognition datasets, for instance, number in the millions, and researchers still complain they are not large enough. In games like chess it is possible to obtain large training sets from starting with all the recorded historical games played between people, and then expanding this synthetically by having your alogorithm play against other versions of itself.

The only place you can get datasets of this size in the financial markets is when you deal with very short-term trading phenemenon that resolve within seconds to minutes. This is probably why many of the very successful quant funds like Renaissance have largely exploited phenemenon at this level. If you are instead training a model to find the best asset-class to invest in over the next six-months, then in the last 40 years of market history you really only have 80 non-overlapping periods. This is far too small of a dataset to do any kind of deep learning on.

More abstractly, what enables computers to learn by "playing against themselves" in games like chess is the fact that the game can be entirely described by a fairly small set of rules that govern how gameplay unfolds and what you are and are not allowed to do. The only variable is what move your opponent will make next. There are no equivalent rules about the financial markets, which may unfold differently every time. This is why the approaches that work for things like asset allocation tend to rely more on simple rules than on complex logic.

Does this mean that AI has no role to play in investing? Absolutely not. Over the coming years it seems inevitable that AI will augment and/or eventually replace many of the things that human stock market analysts at mutual funds and hedge funds do today. A large part of a typical "stock pickers" time is spent evaluating how well a business is performing. If the analyst determines that a company is performing better than the market anticipates ahead of the time that they announce earnings, his firm can profit immensely by buying the shares ahead of an earnings announcement. Hence Wall Street's utter obsession with "earnings estimates."

Many of the analyses that analysts do to determine their estimates could be augmented or replaced by good AI

  • Instead of doing in-store checks on traffic, AI could analyze satelite images and/or geo-location data given off by smartphones to measure real-time traffic trends in retail establishments
  • Instead of listening to earnings calls or presentations by company management, a sophisticated natural language algorithm could analyze everything the company says and detect sentiment changes
  • Changes in consumer sentiment could be detected in real-time from similar social media sentiment analysis

You can use your imagination to easily come up with other scenarios. Rest assured that none of this is new to any Wall Street firm and that the smartest and best firms have been and are actively investing in this kind of technology.

Of course, what none of this will change is the indisputable logic of the zero-sum nature of trying to beat the market by picking stocks. As sophisticated as any firm's AI technology gets, to earn "alpha" (excess return above the market) it will have to be better than the competition's, and the competition also will not rest. In other words, it is highly likely that the stock-picking / hedge-fund world will quickly turn into an AI arms race of Goldman Sachs vs. Bridgewater vs. Renaissance vs. Fidelity, etc. The firms that are truly "best-in-class" at AI may be able to stay ahead of the competition for a bit and earn excess return for it, but the investments required to get there will not be cheap, and the additional investments required to stay at the top of the field may end up consuming a large part of the excess returns that the funds earn for their customers from being there.

So what should you make of all of this as an individual investor? It's pretty simple really. If you are going to play the AI game, you need to be able to win it. If you only have second-class AI, you will lose to the firms making larger technology investments that have first-class AI. Unless you are one of the 50 best data scientists in the world, your odds of beating everyone else at the AI game would rationally seem slim. If winning yourself is not an option, that means entrusting your money to a firm that you believe can. That means correctly identifying that firm and hoping that the excess profits they can generate will be enough to cover the fees they will charge you plus leave some excess leftover to reward you for taking the risk. If you can get access to the next Renaissance Technologies (or for that matter, if you can get access to the current Renaissance Technologies), you should probably do that. But if are reading this article, you probably aren't in that camp.

But that's okay, because the "zero-sum" nature of the markets has an upside -- it gives individual investors a great strategy when faced with games they can't win: don't play. The reason this is a great strategy is that "don't play" does NOT mean "stop investing." It just means "invest in a way you aren't trying to beat top AI firms at their own game." Buying low-cost index funds and concentrating on asset allocation is a great way to do this -- you aren't competing with anyone else on the zero-sum question of whether Target's earnings are going to beat consensus or not, you simply earn the market-average return of all US stocks by buying a few shares of VT, and then focus on the question of how much of your portfolio should be in US stocks vs. other assets.

To put it another way, the good thing about "simpler" rules-based strategies like the IvyVest strategy, or even basic indexing, is that while it doesn't use any AI, it's also pretty AI-proof. As the leading firms fight it out with ever and ever more sophisticated algorithms, so long as you as an individual investor can stay invested, keep your fees low, and avoid the worst bubbles and bear markets with a rules-based dynamic strategy like the one on this website, it's virtually guaranteed that you will end up beating most other investors over time -- even the ones that have rushed headfirst into AI.

Reference 1: Richard Rubin, Margaret Collins, "How an Exclusive Hedge Fund Turbo Charged Its Retirement Plan, Bloomberg, June 16, 2015.


Get our next article delivered to your inbox.

Sign up below and be the first to know about our freshest data-driven thinking on the markets, and investing. We will send you no more than one email a week. This is free.

Ready to start putting this into action?

Take a free two-week trial to IvyVest premium -- our premium subscription service. You'll get access to our rules-based dynamic asset allocation model, tools that will show you exactly what you need to buy in your own discount brokerage account (and when to re-balance) to implement it for yourself, and an insightful monthly newsletter that will keep you on abreast of the most important things going on in the markets. There is no credit card required. Get Started Now!


By Alex Frey