ChatGPT, the artificial intelligence chatbot from OpenAI, has become a global phenomenon, attracting more than 100 million active users in just two months after its launch — a record for the fastest-growing consumer application in history1. The technological models powering ChatGPT, advanced machine learning algorithms, have made rip-roaring waves in a wide range of fields, from protein structure prediction in biology to creator tools in social media. In this piece, we consider how AI and machine learning2 could apply to investment decision-making. It turns out that integrating AI into the financial system requires thoughtfulness, as finance does not neatly meet the criteria of categories that are compatible with AI.
See glossary of terms at the end of the document.
Machine learning is a paradigm shift from traditional science. Unlike the traditional scientific method that involves forming a hypothesis and testing it to draw conclusions, machine learning uses statistical techniques to uncover patterns in data to find answers; no hypothesis or explicit understanding of the system is required. As a result, machine learning can find an open door to questions that are too multifaceted for humans to fully understand, or that are too time-consuming for consumers — or companies — to figure out.
Often, discerning the patterns from data — a key strength of machine learning — is enough to solve many real-life problems. For example, in the fields of natural language processing and image processing, machine learning has been so highly effective that it has fundamentally transformed the way things are done. However, since AI is domain-agnostic at the algorithm level, we can surmise that certain shared characteristics of those systems are conducive to finding answers by pattern detection.
Those characteristics are:
For systems with the above traits, we conclude that if a machine-learning-based model is rigorously trained and validated in a testing environment, it will also perform well with yet-to-be-seen live data.
We considered whether the above traits apply to the financial system, and tested the crucial assumption that machine learning models, if well-trained and validated, will perform similarly when making investment decisions in an out-of-sample, or “real world,” scenario. In our view, several aspects of the financial system make it less apt for the integration of AI:
At bottom, we do not believe that machine learning models, unless carefully designed with investment expertise and executed within a rigorous research discipline, will predict favorable investment decisions with the same accuracy in an out-of-sample scenario. Due to the above challenges, the risk is high that machine learning models may “overfit” investment data, meaning that the model is set to fit too closely to the training data and therefore does not generalize to unseen data. In our view, an entirely data-driven approach — as performed by typical machine learning applications — is unlikely to succeed when tackling investment problems.
To derive tangible benefits from AI in client portfolios, which is, after all, the primary goal of any financial model, the keys are: (1) to understand the conditions required for machine learning to be successful, and (2) to apply the model’s outputs in the right way. For a prudent investor, the calls to integrate AI could prompt a journey to find the right level of trade-off between the traditional modeling culture, where major decisions are made by human analysts, and the data-driven culture of machine learning.
For the financial system, we considered steps that may allow investors to overcome the three challenges (system stability, data availability and quality, and ease of understanding outputs) listed above.
Below, we review a selection of cases in which the Systematic Equity; Investment Solutions Group (ISG); and Fixed Income, Cash, and Currency (FICC) teams have used machine learning to improve investment outcomes in client portfolios.
The availability of unstructured data has increased in recent decades, with over 80% of data available on the internet being unstructured.4 In finance, unstructured textual data such as earnings call transcripts or regulatory filings has been a relatively untapped source of information, but it can offer valuable insights into a company’s operations and business prospects.
The Systematic Equity team employs a suite of factors based on natural language processing (NLP). These factors are designed to capture return-influencing features and insights from unstructured textual data. For earnings call transcripts, for instance, they apply a variety of techniques from traditional linguistic processing to advanced machine learning to analyze a wide range of measures such as the tone and complexity of the language used, the subtle sentiment behind executives’ statements, and the management’s behavior during the call (see Extracting Complex Investment Insights from Earnings Calls). These combined measures proved to assess the overall sentiment and quality of announced results beyond the immediate financial results released in each reporting cycle. Other useful textual data the Systematic Equity team processes includes regulatory filings, patents, job postings, etc.
One of the advanced machine learning techniques used by the Systematic Equity team is a neural-network-based “embedding” that converts the textual data into lower-dimensional numerical vectors, preserving semantic and syntactic information. Once the texts are converted, the companies that discussed similar topics will appear clustered together in the resulting numerical vector space (Figure 1), which can serve as a useful byproduct to identify novel peer groups beyond the traditional sector/industry classifications.
Figure 1: NLP can help group companies discussing similar topics on earnings calls
In the above figure, each point represents and earnings call transcript that was published in May 2018. For each transcript, 300-dimensional embedding was generated using a neural-network-based embedding technique. Embeddings were then projected onto the two-dimensional space for ease of visualization. While most of the companies belonging to the same sector are clustered together, there are also clusters consisting of companies from multiple sectors. One example is Amazon (Consumer Discretionary), placed much closer to Alphabet, Facebook (Communication Services) and Microsoft (Information Technology) than its sector peer Panasonic.
A similar machine learning technique can also be used to identify companies relevant to certain “themes,” such as particular technologies or macroeconomic trends, by processing the filings, earnings call transcripts, and various disclosures from the companies.
Overall, NLP is an ideal way to apply machine learning in finance, as one can train the model using a large corpus of textual data without relying on noisy market data or even borrow pre-trained language models directly from the mature field of NLP machine learning research. And the core language model will remain stable even when its outcomes may need to be used differently in response to changing market conditions.
In tactical asset allocation (TAA) (see Why Institutional Investors Should Consider Tactical Asset Allocation), the relationship between asset prices and underlying factors is often complicated and far from a simple linear dependency. ISG has been a pioneer in what we refer to as regime-driven tactical investing (see State Street Global Advisors Market Regime Indicator: Q422), which is based on assessing a market environment’s “regime,” or level of risk aversion. The ISG team combines the regime indicator models with machine learning techniques such as hidden Markov models (HMMs) to evaluate the relative attractiveness of various assets and forecast total returns.
The ISG team employed HMMs first in 2004 as part of their process to forecast commodities prices, and later in 2010 to tactically allocate between emerging and developed equity markets. HMM remains a technique the ISG team employs, when appropriate. In 2022, they developed a new model for predicting credit spreads5 that employs an HMM. The ISG team plans to deploy it as part of their TAA strategy in 2023.
The ISG team has also been using dynamic linear modeling for situations in which the time variation of asset price to factor relationship is material, but gradual. Such machine learning techniques as the Kalman Filter have also been utilized by the ISG team since 2014 in application to fixed income sector rotation problems and in their hedge fund replication strategies. Finally, in 2022, the ISG team investigated if a decision-tree-based approach is superior to a simple linear model when using macro data changes as factors for tactical equity positioning; early results look promising.
The Systematic Equity team has developed an XGBoost-based method to forecast company fundamentals such as earnings and cash flows, crucial inputs for equity investors in determining the expected return of a company stock. The machine learning method, which allows the modeling of complex relationships among the trailing firm fundamentals, fast-moving market variables, analyst estimates, and macro indicators, proved its clear advantage in terms of accuracy over the plain sell-side analyst estimates across multiple regions. While the method has not been implemented in live portfolios, as overlapping information was already present in other parts of the model, machine-learned company fundamentals can serve as useful inputs in other contexts, providing in-house estimates with better coverage and fewer biases than conventional analyst estimates.
The relatively long time horizon of tactical investing, combined with the relatively small number of available degrees of freedom, makes reducing estimation noise a key focus of our efforts. The ISG and FICC teams made early attempts in employing such machine learning techniques as regularized regressions (such as LASSO, probit, or others) more than a decade ago. More recently, those teams introduced additional techniques, some of which they developed internally. In 2019, the ISG team incorporated random correlation matrix cleaning into our portfolio construction process. In 2020, they suggested an adaptive enhancement to a classical risk parity strategy6 that can be viewed as a machine learning approach. In 2021, they conceptualized on how to incorporate fuzzy mathematics — a well-known AI technique — into portfolio optimization as an antidote to the estimation noise7. Currently (March 2023), the ISG team is working on a risk model that uses PAM clustering as a way to intelligently reduce the complexity of the emerging market bond universe.
Much ink has been spilled about AI and its applications across multiple categories. For financial systems, AI can be useful in offering insights from sources that are infrequently tapped, modeling asset prices and factors with nonlinear relationships, or reducing estimation noise, among other uses. What’s important, however, is understanding the contexts in which machine learning outcomes are the most helpful and reliable. In our view, machine learning is best when combined with the knowledge of strong investment teams that can control the structures, contexts, and training and determine the questions that AI will address. When used thoughtfully, AI can be a powerful investment tool to unlock complex, previously untapped information from a wide range of variables.