09.01.2023
AI in Investing
For decades, Artificial Intelligence (AI) has excelled at complex mathematical tasks and strategic reasoning. In fact, early versions programmed to play Chess and Go became so good at the games that they were almost unbeatable, while later versions became unassailable. Today, AI’s capabilities have grown well beyond strategic games thanks to more powerful machine learning technology—and those new capabilities are about to transform the world of investing.
What Is Artificial Intelligence?
The term has been broadly applied to a wide range of technologies, but AI generally refers to a machine that is capable of human-like reasoning and autonomous learning. Some recent examples of the different kinds of AI technology include:
- Expert systems like autopilot for commercial aircraft are artificial intelligence programs that use multiple techniques like rule-based systems, fuzzy logic, and neural networks to help pilots make decisions and solve problems. They're used in aviation safety to analyze data and warn about potential risks, such as avoiding runway issues. These systems also assist in flight planning by considering factors like weather and air traffic to find the best route. Many modern aircraft spend much of the flight with AI fully in control.
- Natural language processing, like ChatGPT, uses large language models to analyze things like the meaning, intent, and sentiment behind text in order to answer questions, summarize reports, and assist with other writing and reasoning tasks.
- Neural networks like the AI-powered speech recognition software called Whisper which uses a neural network to accurately distinguish english speech from background noise or other languages being spoken, even across different accents or speakers using industry-specific technical language.
Combinations of these types of AI technologies have been used in Finance for over two decades mainly in areas such as fraud detection for credit card transactions, assessing loan applications and network intrusion detection using pattern recognition.
More recently, AI has been used in banking apps to make product recommendations or to power chatbots for customer services and issue support. AI has also been used as a stock recommender as part of a portfolio management objective in some trading apps.
Each of these different technologies work a little differently but all of them dramatically advance our ability to recognize patterns and create rules around those patterns. These patterns and rules can, in turn, help us make better decisions. In other words, AI represents a collection of technologies and abilities that have the potential to transform society by offering novel forms of value and changing the way we interact with technology.
A timeline of how AI has already changed the way we invest
Advanced AI based on neural networks and machine learning is not new. In fact, it can trace its origins back more than 70 years to early computer scientists and mathematicians developing the first computers, including early prototypes of machines with problem-solving and language interpretation capabilities.
AI isn’t even really new to investing, as computers have been used in trading and investing for decades and in some sense modern AI in investing is a continuity of these early developments. Let’s take a look at some of the key historical touchpoints of how investors have been using computer programming to automate trading and even aid in making trading decisions.
Programmatic Trading in the 1980s
In the early 1980s, fueled by advancements in technology and financial innovations such as derivatives, institutional investors began using computer programs to execute trades based on predefined rules and algorithms. This helped them complete large trades faster and more efficiently than they could before.
At the time, these algorithms were relatively simple and were primarily used for so-called index arbitrage, which involves trying to profit from discrepancies between the price of a stock index – like the S&P 500 – and that of the stocks it’s composed of. Later on, institutional players found value in computer programs managing order execution for large order entry or exits.
High Frequency Trading in the 2000s
In 2002, when the New York Stock Exchange introduced a fully automated trading system, programmatic trading gave way to an even more sophisticated form of automation with much more advanced technology: High-Frequency Trading (HFT).
HFT uses computer programs to analyze market data and execute trades at extremely high speeds. Unlike programmatic traders that bought and sold baskets of securities over time to take advantage of an arbitrage opportunity, high-frequency traders use powerful computers and high-speed networks to analyze market data and execute trades at lightning-fast speeds. High-frequency traders can conduct trades in thousandths of a second, compared with the several seconds it took traders in the 80s.
On institutional trading desks, the variety of operations that were automated also expanded from simple and routine trades to medium-sized trades and then to increasingly sophisticated order types. The growth of automated trading algorithms continues to accelerate the evolution of financial markets. As trading desks grow increasingly automated, the role of human traders has evolved and many were lost.
These early examples of computerization on wall street or specialized HFT algorithms were not AI, but I would argue that there is a continuation from these early technologies as they helped lay the groundwork for how modern AI is being deployed in investing today.
How the latest generation of AI is making fully autonomous investing a reality
Today, machine learning and other AI technologies have come so far that they are making possible next-generation capabilities like fully autonomous trading and investment vehicles that investors in the past could only dream about. The ability to augment human intelligence and decision-making is empowering investment managers to create everything from AI-informed funds all the way to fully automated AI-directed funds.
Recent breakthroughs in generative AI have seen the S&P 500 index soar even in an otherwise rocky market, thanks to the massive gains from large tech stocks that have buoyed the index. Likewise, thematic funds built around AI-related stocks and other funds with exposure to those stocks have been especially strong performers in 2023 as that same AI technology promises a new era of growth for countless other industries outside of tech.
Given the vast abilities of AI, like data-crunching and analysis, we have seen the emergence of a number of funds that are still human-directed but use AI to research, pick stocks, conduct sentiment analysis, and perform an increasing number of other analytic and decision-making tasks at a scale that were once impractical for machines to do.
Now, we are seeing new entrants into the investing space that are using completely autonomous AI-directed funds. After designing and training the AI to follow a particular trading strategy, the fully autonomous AI does all the work of scanning the market for trades and identifying the entry and exit signals. In this case, the only job of the human fund manager is to serve as human oversight to manage and contain risk.
What makes today’s AI different from the programmatic and high frequency trading of the past?
While using machines to automate trading is nothing new, the advanced neural networks and machine learning that drives this generation of AI is. It is that powerful pattern recognition and self-adjusting technology that is making AI-informed and fully AI-directed funds possible.
A neural network is a large web of connected processors that work together to achieve complex tasks. The design was inspired by the way neurons connect in the brain. When data is fed into the network, each processing node performs its specific analytic function before sending the data on to the next node. After the input has worked its way through the network, guided by the strength of the connections, the neural network is able to find the appropriate output that meets some optimization function, providing a near best answer possible within that trained network.
To understand what makes this neural network model so powerful, consider the self-driving car. The neural network powering the car receives inputs from various sensors located all over the vehicle. Those inputs include things like how fast the car is going, what kind of surface the car is driving on, where other objects are in relation to the car, and so on. After those inputs are fed through the neural network, the output could be decisions on how to steer, whether to speed up or slow down, and when to break.
To train the vehicle, developers feed it massive quantities of data, like images and videos of humans driving, busy roads, traffic lights, street signs and other relevant information. Using that training data, the neural network strengthens the relevant connections and eventually learns which inputs correspond to which actions so that it can drive a vehicle by itself. That includes everything from plotting the best route to knowing how to respond to different traffic signals and signs to distinguishing between a statue and a live human being.
Neural networks have been around since the 1950s, but only in the last 15 or 20 years have they started to make real progress. The change of fortunes is due to the emergence of new mathematical frameworks as well as the significant increase in both the amount of data we produce and the amount of computer power available to process that data.
What Neural networks do extremely well is recognizing patterns in data and identifying the rules or associations between different input sets. Traditional computer programming involves a human writing step-by-step instructions that tell a computer what to do. With neural networks and machine learning, on the other hand, computers can learn to structure data without direction.
Instead of a human telling the computer what correlations to make between inputs, it is the computer that finds those correlations itself and represents these patterns in that network of nodes loosely based on the brain. These representations can then be used to predict future outcomes based on new inputs that it has never encountered before.
It is this ability to apply its training to new data as well as learn how to refine its pattern recognition and predictive functions based on that new data that is proving to be so transformative in the investing space. By training these neural networks on vast amounts of historical market data these systems can be effective at predicting future risk and returns based on current market conditions within a limited scope — similar to what a stock portfolio manager does over a narrow time horizon.
Fully autonomous AI still needs human guardrails
To use AI responsibly in investing, investment managers still need to maintain tight guard rails around the internal decision making of AI. While the true value in AI is in identifying trends in vast quantities of data and making judgments based on that at supernatural speed, these decisions need to be limited and constrained to meet specific objectives within hard boundaries.
These clear objectives and boundaries prevent the AI from taking on inappropriate risk. For example, an investment manager might instruct an AI to maintain a certain level of diversification to avoid concentrating in one asset. Just like human-directed trading strategies, AI will be governed by a clear investment policy. Unlike humans, though, AI will not be prone to making mistakes that break those hard rules.
There are still limits to what AI can do
As AI is starting to take center stage as an Investment tool we must also recognize that it has important limitations and cannot do everything well. AI is a very different type of intelligence than human intelligence. Without embodiment, without a world view or an understanding of the real world, whether that’s physical objects or the macroeconomic and geopolitical circumstances on the world stage, AI is simply too limited in scope to manage multidimensional contextual problems well.
However, it is going to be vastly superior in the scale of data it can consume and process for specific indicators compared to humans. For instance monitoring hundreds or thousands of stocks for specific patterns in behavior is something that AI can do without sleeping.
Programmed together with objective and hard limitations on its scope and scale, AI will be able to run successful limited trading applications. This is already happening as we have a number of completely AI-driven ETF products on the market and we expect this to increase rapidly in the coming years. As its ability to learn from its past decisions allows it to continuously hone its pattern recognition and predictive capabilities, the AI powering these funds will keep getting better, making the emerging AI-driven ETF space one every investor should watch.