Over the past few years, the adoption of Artificial Intelligence (AI) and Machine Learning has been a key driver of innovation in many industries. Tasks like facial recognition or assisted medical imaging ― unthinkable just a few years ago ― are now possible thanks to models that learn from data to understand and predict the behavior of complex systems.
Similarly, investing ― and its vast toolkit of quantitative techniques ― is greatly benefiting from the introduction of AI. Indeed, the new landscape is taking shape at a time of increasing complexity and competition in financial markets, making the quest to find information hidden in the noise more important than ever. Yet, as we will see, AI and machine learning mark the pivotal transition from human (and statistical) reasoning to a new era of assisted decision-making.
They are the new forces shaping what investing is going to look like: a tech challenge. Their value ― instead of revolutionary ― should be considered evolutionary, i.e. sophisticated tools to tackle known problems.
Introduction
Recently, the application of AI to investing has become a hot topic in financial literature. Behind that, the availability of big data and the ever-increasing computing power are making the decades-long research in the field of AI finally deliver on its promise of performing human-like tasks. With AI, the financial industry is now witnessing the evolution of a long tradition that ― since the early 50s ― has leveraged quantitative techniques for a variety of purposes, from securities’ analysis to portfolio management. Yet, the evolutionary pressure comes as no surprise. The increasing complexity has required reframing investing in terms of an unbiased and adaptive approach to financial markets. Machine learning models ― the subset of AI involved with learning from data ― are gaining traction as new tools to support managers in their investment decision-making process.
Start integrating AI into your investment process with the most advanced no-code solution.
Book A Custom Demo
Decoding Complexity, One Example at a Time
New AI models differ markedly from traditional models in how they address problem-solving.
Specifically, instead of explicitly programming a machine to perform a sequence of consecutive tasks ― as happens for traditional algorithms ― machine learning algorithms are designed to learn the solution from vast sets of data. This makes them suited to problems that are not only complex by nature but entail high-dimensional datasets and evolving dynamics.
The following table summarizes the main differences between the two approaches:
In this sense, if the solution to a problem is known beforehand, a traditional algorithm and a machine learning algorithm would return the same solution in the long run. Indeed, if we were to solve a known problem with machine learning, the solution would eventually be detected as more data is fed into the system. From this an important remark: the advantage of AI models is to tackle problems without the ex-ante knowledge of the relevant variables necessary to its solution.
This is particularly true for problems that have always demanded an intelligent way of making sense of the magnitude of data collected, like space exploration.
“We won’t need eyes to see”: NASA is Exploring Space with AI
In the words of Dr. Weir from the famous 1997 movie Event Horizon, “Where we are going, we won’t need eyes to see”, AI is now used at the frontier of space exploration to discover new celestial objects hidden in the tons of data collected from space. For instance, in the last few years, we have seen thousands of earth-size planets outside our solar system (i.e. exoplanets) being discovered using data from the Kepler telescope launched by NASA in 2009.
Just to get a sense of the magnitude, the four-year dataset researchers have used consists of about 2x1015 (i.e. 2 quadrillions) possible orbits of planets to sift through. But how can data be used to detect a planet in the first place? In the case of Kepler’s data, the telescope recorded the amount of light received from around 100’000 stars in its field of view.
If a planet were to orbit around one of them, the star’s brightness would decrease slightly at each overlap. However, best-in-class traditional methods such as radial velocity and transit photometry ― although very powerful ― suffer the drawback of missing many of the weakest signals and still rely on human judgement for their verification. This is where AI enters the stage.
Exhibit 2 - Main differences between a traditional and a machine learning algorithm.
As AI models learn from data the relevant rules to solve a problem, astronomers have leveraged this feature to train an algorithm (i.e. a Deep Neural Network) that identifies new celestial objects learning from Kepler’s previous discoveries, almost 15’000 labeled examples.
The trained model that was subsequently deployed was able to detect many of the weakest signals researchers had previously missed. A group of low-intensity signals eventually led to the discovery of KOI-351i, better known as Kepler-90i ― a super-Earth exoplanet located 2’545 light-years from Earth.
Behind the Scenes: How do Machines Learn?
But apart from the gigantic leap in performance, the real technological breakthrough in AI models lies in how algorithms can be designed to learn from data. As we dig deep into the foundations of human learning, we understand our knowledge building process is a virtuous circle of information gathering and a constant feedback loop received from the outside environment. Specifically, learning from examples and experience is how humans evolved to solve problems efficiently and quickly adapt to new information.
On this premise, developing a new generation of algorithms able to mimic this process has demanded rethinking completely the purpose of existing models ― in this sense, not being precisely wrong, but approximately right.
Data-driven learning comes as the wise mix between high-quality input data, the appropriate mathematical model together with a way to update and monitor knowledge as new information becomes available ― the learning method. Indeed, one of the most famous machine learning models out there ― the Neural Network ― has been designed to mimic the architecture of the human brain neurons, that is, a series of self-reinforcing connections between inputs and outputs, as shown by the following images.
Exhibit 4 - Stylized form of a human neuron (left) and the structure of an Artificial Neural Neuron (right).
In this sense, as a biological neuron processes the information coming from dendrites into an output signal down the axon, a Neural Network weights initial input data through layers of mathematical operations in an output response. This mechanism allows the model to discover the hidden relationship behind input and output data using a technique called backpropagation as the driver of the learning process. Specifically ― as the network is called to make forecasts weighting input data ― it seeks by trial and error to better balance initial information.
Eventually, learning from a vast collection of examples of what is correct and what is not, it will gradually adjust its weights in order to reduce the error rate to the minimum. In this way, the network is said to learn from experience. Ultimately, learning involves searching for the optimal way to connect the dots to solve a problem.
Similar to a human brain, a Neural Network expands its knowledge as more data is collected and this turns into a better understanding of the dynamics we are analysing.
One of the reasons this particular field of research has witnessed explosive developments is the (recent) availability of big data sets. State-of-the-art technology (as seen in space exploration) focuses on training more complex network architectures to model problems of higher dimensionality using so-called Deep Neural Networks ― Neural Networks with many layers between input and output.
Exhibit 5 - The structure of a Simple (left) and Deep (right) Neural Network.
Similar Yet Different: Why Financial Markets are Unique
The applications of this new technology are indeed ubiquitous. As a matter of fact, AI does not only take problem-solving one step further with a set of tools to extract information from big data, but it is quickly becoming a strategic tool to assist decision-makers in a complex and ever-evolving environment.
The investing industry ― and financial markets as a whole ― has been an early adopter of this new technology, seizing the opportunity to improve the efficiency of investments.
However, the tools of AI are better used for very specific purposes within the investment environment, rather than being one-stop solutions for better performance. Instead, they are best used when there are both deep expertise and a thorough understanding of the reasons why it could be beneficial to specific problems. Yet, it is undeniable the extent to which AI models have a clear competitive advantage over humans when it comes to calculating and measuring phenomena.
Indeed ― similarly to astronomy ―markets speak to investors with tons of numbers and data. Nevertheless, real-world investing is different. As a matter of fact, if on the one hand financial variables change over time (e.g. think of the big changes in correlation between stocks and bonds in the last 10 years), on the other hand we also see financial data to be characterized by a great amount of noise that makes valuable information very hard to extract.
In this regard, as financial markets are complex and ever-evolving systems, they make the ideal setting for tools and technologies to extract information from data. On top of that, the economics that drive financial markets make them an intrinsically dynamic system, where every bit of new valuable information is quickly used to invest and arbitraged away (i.e. exploited by the trading of investors).
Metaphorically, we could say that ― in financial markets ― a star begins dimming as soon it is discovered.
Reframing Portfolio Management
The benefits of AI stand at the crossroad between a deeper understanding of the fundamental drivers and a scientific approach that avoids human bias to enter the investment process. Indeed, as financial markets have become a more complex environment, a new technology-driven field of competition has emerged ― transforming investing into a tech challenge.
In this sense, portfolio management (i.e. the investing subset that involves portfolio decisions) has long adopted quantitative tools for a variety of purposes such as securities’ analysis and portfolio construction.
To this day, investing themes such as factor investing or the efficient portfolio methodology rely heavily on the use of mathematical models to analyze financial data. That is the reason why applying machine learning models has the potential of giving investors an edge in terms of a better signal extraction that leverages the increasing volumes of available data. AI ― instead of revolutionary ― comes as the result of a series of technological innovations that are reframing portfolio management.
Current state-of-the-art research is focused on the application of machine learning models in the following areas:
Securities’ Analysis
One area in which machine learning is finding vast application is securities’ analysis, that is, using available information such as prices, volumes, balance sheet figures, news and alternative sources of data to better understand the securities to be included in a portfolio.
Traditionally, the output of this process is a judgment about the under/overvaluation of a security relative to a group of its peers or, alternatively, an analysis to find out the drivers of investment performance in terms of risk exposures and investment factors.
However, with machine learning models, its area has been expanded to more complex tasks as, for example, the continuous detection of market inefficiencies ― exploitable anomalies in price generated from investors’ behavior.
Portfolio Construction
The new tools of AI have also proven to be efficient mathematical techniques to address the issues connected to dealing with highly complex datasets.
Quantitative portfolio construction ―relying on a model to perform tasks such as Asset Allocation ― has its roots in the Markowitz and Black & Litterman models that introduced the ideas of risk-diversification and a forward-looking approach to the evolution of securities’ returns.
In this sense, AI models ― especially Neural Networks ― are proving to be useful tools in supporting existing models to learn the continuous evolution of markets and cut through the complexity of high-dimensional data.
Risk Management
Finally, models that learn from data how to adapt to evolving markets have given decision-makers a new toolkit to manage portfolio risk more efficiently. For instance, machine learning is currently being used to generate real-time risk forecasts, going beyond the ex-post approach of traditional risk management measures. Yet, a major shift has been witnessed in portfolio diversification ― the mechanism that allows gains and losses to balance each other and reduce the overall volatility of an investment.
In this sense, improved risk management translates into efficient portfolios that better protect capital over time. Indeed, machine learning has allowed correlation analysis ― the source of diversification ― to be performed not only on a larger scale but to adapt continuously as financial markets change over time.
Summing Up: An Evolution, Not a Revolution
In an era of increasing complexity and technological breakthroughs, AI addresses the demand for a data-driven and adaptive approach to financial markets. That is why AI and machine learning mark the natural evolution of the established approach to quantitative investing.
Yet, the gains they bring are not revolutionary but rather evolutionary, as new tools to address specific aspects of complex problems.
As a result of this, investing has become a tech challenge in which technology is allowing to uncover a layer of hidden information that can be only harnessed through a deeper analysis of data.
