AB IQ Artificial Intelligence (AI) Systematic Tech and Innovation Alternatives Equities Fixed Income Blog

Deploying AI in Investment Applications

Three Case Studies

29 September 2023

7 min read

Bernd Wuebben| Director—Systematic Investing and Quantitative Research

Jonathan Berkow| Director of Quantitative Research and Data Science—Equities

There are different ways to apply AI to portfolio management processes. We discuss three distinct approaches in fixed income, hedge funds and equities.

Artificial intelligence (AI) offers investors specialized tools to cope with the vast, complex and rapidly changing financial markets, but its applications differ from strategy to strategy. In these case studies, we show the diverse ways AI is being applied to enhance the investing process and pursue better client outcomes across asset classes.

Case Study 1: Empowering Systematic Investing in Credit Markets

AI is quickly becoming an indispensable tool for active credit managers, helping them navigate an ocean of constantly changing data to analyze opportunities, making it an excellent fit for credit investing and systematic fixed-income approaches.

Think of systematic bond investing as a three-step process: identify which bonds are attractive and unattractive using an objective ranking system; optimize those findings to create a portfolio; then realize that portfolio through skillful implementation. AI can improve the first and third steps by helping decide what to buy and sell, estimating missing data and finding pockets of attractively priced liquidity.

Generating Fixed-Income Alpha with AI

Systematic fixed-income strategies seek to outperform bond-market benchmarks principally through individual security selection. They aim to rank the securities in a benchmark to identify outperformance (alpha) potential, using an array of predictive factors based on various types of data or signals, such as valuation, momentum or sentiment.

Machine learning (ML) methods (Display) can improve analytics across multiple valuation factors to find new signals and make existing signals more effective, enhancing systematic managers’ ability to rank securities. AI can, for example, derive valuation scores at both the individual bond level and the issuer level, creating different lenses through which to identify bond-price anomalies. AI also brings a new dimension to the probability of default analysis, improving on rigid academic models to deploy a greater number of analytics more empirically across a broader dataset.

AI Glossary: Basic Terms

Definitions provided for key AI terms including: machine learning, deep learning, natural language processing, transformer-based models and large language models.

Source: IBM, NVIDIA, Oracle and AllianceBernstein (AB)

The breadth that AI brings to research is an important factor in natural language processing (NLP) analysis, too. A team of human analysts can digest only a fraction of the available data from a glut of company reports and is unable to analyze the information embedded in words and phrases consistently across different companies. AI’s transformer-based language models can handle the full dataset with complete consistency, creating breadth that improves the quality of active return streams for systematic strategies.

Beyond individual security selection, which we believe is the main advantage of systematic fixed-income portfolios, traditional active management uses other approaches. These include sector and industry tilts and timing strategies based on interest-rate risk (duration) and credit risk. In our view, these strategies can also benefit from AI technologies. We find that ensemble-based ML models are particularly beneficial, giving rise to risk indicators and beta timers with effective predictive power. Newer methods such as topological data analysis, dynamic time warping and signatures are also potentially promising sources of alpha.

Enhancing Implementation and Bond Trading with AI

Processing bond prices might just be the most boring job in the world. Combing through reams of closing prices and trying to work out the correct values for missing data is a repetitive, laborious and error-prone task. Not for AI. ML can identify price patterns to impute missing data quickly and reliably—and across a complete dataset. Similarly, AI can impute missing liquidity data from trading patterns in a wide range of related securities.

What does that mean in practical terms? For systematic bond strategies that trade corporate credit, the limited liquidity and tradability of certain bonds makes it challenging to capture and understand intraday credit market movements and to extract accurate intraday market signals. AI provides a highly effective solution.

At AB, we leverage advanced ML models to estimate real-time prices for corporate bonds with stale or missing prices, using real-time prices of liquid peers and other relevant market variables. This model-assisted price approach provides us with a more complete view of the trading universe and serves as a foundation for our intraday market signal research.

Forecasting future liquidity profiles across the trading universe is essential for day-to-day credit trading. Here, ML applications can help identify nonlinear relationships between bonds’ characteristics and their future liquidity profiles. Understanding the probability of trading a given bond in a required size with an acceptable spread is crucial to implementing a systematic portfolio successfully. In this area, interpretable machine learning models can deliver strong results, in our view.

Case Study 2: How to Build an AI-Driven Hedge Fund

Imagine a portfolio manager who has studied at the world’s best universities, has a photographic memory and a profound understanding of macroeconomics and geopolitics. She’s an avid news consumer and possesses staggering amounts of information about stocks, bonds and currencies. Her brain flawlessly scans the details of her base knowledge to identify the most attractive assets that could be affected by the ripple effects of current events.

Humanlike Decision-Making at Superhuman Speed and Scale

Of course, this superhuman investor doesn’t exist. The human brain isn’t physically capable of holding multiple gigabytes of diverse information, structured and mapped, and ready to be explored instantly when triggered by a keyword. But we believe that AI-driven models and processes can potentially replicate the human decision-making process at machine capacity, making better investment decisions unhindered by behavioral biases.

To uncover opportunities, the AI investing brain must search beyond traditional data sources and quantitative strategies (Display). The human mind could be led astray by mistaking correlations for cause-and-effect relationships. In contrast, an AI investing brain should capture significant events and persistent causality, by identifying how the chain of potential contagion might develop.

Training an AI Investing Brain

Graphic depicts how an AI-investing brain can be trained using unbiased and credible sources of information, with relevance mapping based on graph theory.

*Topics include but are not limited to finance, geopolitics, technology and macroeconomics. Sources include books, research articles and opinion pieces from Bank of Japan, European Central Bank, International Monetary Fund, World Bank, World Trade Organization, etc.
†In the context of NLP, graph theory refers to the application of graph-based models and algorithms to analyze and represent linguistic structures. It involves representing language data as graphs, where nodes represent linguistic units (such as words or phrases) and edges capture relationships between them (such as syntactic or semantic connections). Graph theory in NLP allows for the exploration of relationships, extraction of meaningful insights, and development of algorithms for knowledge graph construction.
Source: AB

That means focusing on impactful themes and finding financial instruments that might be affected in unobvious ways. In other words, the AI investing brain should be able to determine which stocks, bonds and currencies have not yet been priced into the effects of events. Validating the timing and quantifying the expected alpha requires highly nonlinear, multidimensional and cross-asset ML models.

NLP with a Difference

Many investment teams use NLP technology to process texts and create sentiment signals. But since these daily signals are quickly priced into securities, they tend to be correlated with momentum strategies, in our view. For an AI investing brain to find alpha potential, it must identify more persistent patterns, with longer-term effectiveness that leads to alpha potential uncorrelated from other asset classes.

That requires training a model that covers a specialized yet diverse set of topics to identify subjects, including geopolitics, technology, finance and macroeconomics, that could impact asset prices. The AI brain is fed a regular diet of information from credible sources that we believe are unbiased—from the International Monetary Fund to the World Trade Organization and global central banks. It is constantly evolving and now has tens of millions of nodes (i.e., words and combinations of words) in its knowledge graph.

Utilizing the key themes, this evolving knowledge graph taps into current events in real time. It draws on a select group of reliable media sources and ranks the news based on what is more likely to move financial markets in a persistent way versus short-lived hype. The top-ranked, potentially most influential news items are then used to identify the closest financial assets to invest in the theme.

The Matchmaking Algorithm: Identifying Assets with Alpha Potential

The next step is to connect the dots between influential themes and investment opportunities by gauging the potential impact of each identified theme on financial assets across a broad universe, including stocks, equity indices, government bond indices, commodities and currencies.

When a human investor identifies a theme, it can’t always see the first-, second- and third-order effects. For example, tighter carbon regulations would obviously affect carmakers and key raw materials used to reduce emissions. But the ripple effect goes much further and would be harder to see; it could extend from substitutes for palladium (used in catalytic convertors in vehicle exhausts), to the electric vehicle value chain, to other industries that depend on lithium from technology companies producing chips and hardware, to renewable energy storage systems.

An ML model can extract patterns and quantify multi-asset interdependencies to help forecast returns by providing the direction, long or short, and size for each asset. This model estimates the expected alpha for each security identified as a candidate from a relevant theme. Of course, not every security identified should be selected to be added to the portfolio. Quantitative tools can also help construct a strategy that aims to maximize the alpha potential of short-listed securities while applying multiple levels of risk control.

Case Study 3: Sharpening Equity Research and Portfolio Processes

Equity portfolios that are rooted in fundamental processes might not seem the most obvious candidates for the deployment of AI technology. After all, stock pickers use deep research, sector and company expertise, and most importantly, human judgement to identify alpha opportunities.

We don’t think AI will replace human analysts for active equity portfolios anytime soon. Yet the need for sophisticated tools to manage risk and create portfolios with strong risk-adjusted return potential has become more important than ever. In our view, AI tools can help improve equity investing processes in three main ways: efficiency, managing risk/portfolio construction and generating signals.

Efficiency: Saving Time on the Boring Stuff

For analysts to be most productive, they must focus on asking the big questions with the best information. Yet in the past, collecting the best information in a data-saturated world was labor intensive and flawed. News reports, official filings and earnings calls generate an endless stream of information. This forced investment teams to focus deep research on a limited number of higher-priority holdings or strong investment candidates.

Now, ChatGPT can be used to summarize transcripts of earnings calls and events that analysts wouldn’t normally be able to attend. When implemented correctly, it’s like the analyst has 10,000 interns to help accelerate the discovery of new information. Internal chatbots can be used to help analysts become more efficient at finding information from an immense pool of notes and reports; chatbots can also be used to help quantitative analysts write code, speeding up important data discovery processes.

Managing Portfolio Risks and Construction

Risk-management portfolio construction is another area where AI can be very helpful.

Cluster analysis is a form of AI that’s been used in our risk-management arsenal for several years. It aims to detect unknown risks that may be lurking in unseen market patterns, which may go undetected by standard risk models. This sophisticated ML technique segments stocks into groups whose returns have been moving closely together over a defined period. For example, it can help separate groups of stocks that relate to a new sub-theme or risk, which can guide the portfolio manager to hedge the unintended positions.

In addition to raw data like stock returns, unstructured text can be used to discover relationships between companies and trends. NLP and word embeddings allow us to source thematic stocks or identify portfolio risks by systematically analyzing materials, including company transcripts, 10-K and 10-Q forms, and Wikipedia data. This provides a scalable way to find stocks that are related to broad themes such as demographics and aging, AI or urbanization, which could be important drivers of risk and return.

Generating Sentiment Signals

AI can also support equity research by generating signals. We believe that these shouldn’t be used in isolation, but they can be effective when paired with fundamental analysis.

Using large language models (LLMs) such as BERT and finBERT, we have long used NLP tools to generate signals such as positive or negative sentiment from CEOs. These tools can also be used to estimate a Gunning fog index, a gauge of the complexity of words and sentences, which may indicate that a company is trying to confuse investors. LLMs are much better than traditional tools at picking up these nuances.

Another set of NLP signals drills into words embedded in corporate filings to identify large changes in real time. By systematically and comprehensively finding sections that have changed, we may be able to discover potential risks earlier than others. Fundamental analysts can then review the 10-Ks of companies flagged to validate the risks detected.

When strategically deployed, AI can help active equity investors make better investment decisions. However, we believe that AI is a tool that must be used in conjunction with human expertise and judgment. While AI can help us process vast amounts of data quickly and efficiently, we believe that for long-term equity investors, it cannot replace the insights and skill that come from years of industry experience.

The views expressed herein do not constitute research, investment advice or trade recommendations, and do not necessarily represent the views of all AB portfolio-management teams and are subject to change over time.

About the Authors

Bernd Wuebben

Bernd Wuebben oversees the Quantitative Research Group for Fixed Income, where he is responsible for developing the quantitative models and methods that are an important aspect of the firm's investment process. Prior to joining AB, Wuebben worked from 2008 to 2018 as a quantitative investment portfolio manager, managing portfolios that invested in liquid traded assets across all major geographic locations and asset classes. His quantitative investment process made heavy use of big data and machine learning. Prior to this, Wuebben worked for 10 years as a macro and relative-value strategist as well as a proprietary trader for major investment banks including Morgan Stanley, Bear Stearns and Deutsche Bank, where he started his career in 1998 as a fixed-income strategist. He holds a BS in mathematics and a BS in physics from Heidelberg University, Germany, and an MS in mathematics and an MS in computer science from Cornell University. Wuebben also passed the doctoral exam in mathematics at Cornell University, where his research focused on the mathematical aspects of superstring theory. Location: New York

Jonathan Berkow

Jonathan Berkow is a Senior Vice President and the Director of Quantitative Research and Data Science in the Equities division at AB. He leads the research and adoption of alternative data in equity research and systematic strategies. Prior to joining the firm in 2018, Berkow was a systematic portfolio manager and researcher at hedge funds Element Capital Management and Kepos Capital. He started his career at Goldman Sachs Asset Management, where he managed quantitative research and was a portfolio manager for global equity portfolios. Berkow's research has spanned equities and macro asset classes. He holds a BS in economics from the Massachusetts Institute of Technology. Location: New York