The AI revolution in retirement

Integrating AI as an investment insight within retirement portfolios

Artificial Intelligence (“AI”) is poised to transform nearly every industry, and that includes financial services and how people can reach their retirement goals. With just a little imagination, the opportunities to integrate AI into our retirement ecosystem seem limitless. To help provide a bit of focus, we highlight three key points for plan fiduciaries to consider as you begin to integrate AI into retirement portfolios and investment menus:

AI as an alpha insight
AI’s potential impact on our ability to save for retirement
AI as a differentiator for your retirement plan

But first…

Context: The breakthrough of AI

AI, or the simulation of human intelligence by machines, has been evolving for decades. Generative AI is the latest breakthrough category in the space, garnering attention for its ability to create original ideas and content. Near the end of 2022, OpenAI released its AI platform, ChatGPT, to the public. ChatGPT uses advanced language technology to create large human-like text outputs—bringing the most recent advancements in generative AI to the masses. ChatGPT exploded in popularity at the fastest pace of any online application, reaching one million users in just five days (Figure 1).

/blk-inst-c-assets/documents/charts/ai-revolution-in-retirement-1.csv bar-chart column-simple false

Fig 1: ChatGPT reached one million users faster than any app

Amount of time to reach one million users

Source: Statista, with data from company announcements via Business Insider/LinkedIn, as of January 24, 2023.

So what is the technology underpinning this disruptive platform? ChatGPT is a large language model (“LLM”) based on generative pre-trained transformer (“GPT”) technology. LLMs are trained using massive amounts of data sourced from websites, books, academic publications, and other public datasets. These models are trained to predict the next word in a text given previous context, and in that process they acquire linguistic skills, world knowledge, as well as basic reasoning skills.

Coincidentally, on the same day that Chat GPT was released to the public, BlackRock’s Systematic investment platform approved an investment insight that leverages the same transformer technology powering LLMs. Let’s explore how we use these models and other big data insights to enhance our investment capabilities.

Paragraph-3,Content Block-1,Paragraph-4,Charts-2,Paragraph-5

Paragraph-6

Paragraph-7

1. AI as an alpha insight

The practical implications of integrating AI into retirement portfolios include the use of AI as an alpha insight—a potential game-changer in asset management. The approach goes beyond investing in companies driving the AI revolution and includes leveraging AI to extract differentiated insights in a competitive landscape.

By identifying unique drivers of outperformance or underperformance, AI becomes an invaluable resource kit for portfolio managers that allows for a more sophisticated understanding of market dynamics, potentially leading to more tailored and resilient portfolios capable of adapting to varying conditions.

Geospatial research serves as a compelling case study. By analyzing vast datasets, including satellite imagery and labor mobility data, AI can extract early insights on economic activities across regions, which can be used to inform macro (e.g., regional) and micro (e.g., company level) tilts in our portfolios.

AI can also be used to analyze traditional sources of data more efficiently. For decades, we’ve been applying natural language processing (“NLP”) techniques across a wide range of text sources including broker analyst reports, corporate earnings calls, regulatory filings, and online news articles. When analyzed at scale, each individual insight can be combined into an aggregate view that helps inform our return forecasts. The more effectively we’re able to extract and understand these insights, the more of an investment edge they may be able to provide.

Early investment signals used sentiment analysis, tracking the number of positive and negative words included in a document, and assigning an overall sentiment score based on word counts. While these signals proved effective, they weren’t designed to factor in nuances like sentence structure and semantics that can impact the meaning of text.

New research innovations have improved the granularity of text analysis over time. For example, advancements in machine learning helped to determine the most relevant words to track based on the type of text input.

Utilizing transformer-based LLMs to improve our investment predictions.

Today, our approach has evolved to utilize transformer-based LLMs (just like ChatGPT). Transformers are a type of neural-network architecture that can process long sequences of elements (like words in a sentence), accounting for the relationship between each individual word with other words and focusing on the most important points (Figure 2).

Figure 2: The transformer architecture allows for a more accurate analysis of text by considering the interactions between words in a sentence and identifying the most significant relationships

hypothetical example of text analysis using transformer architecture

Figure 2 is a hypothetical example of how the attention mechanism in transformer models are often illustrated. The right-side shows a particular word, in this case “company” and how a model considers the association of that word with all the other words in the sentence which are listed again on the left-hand side. In this case, the model is recognizing significant importance between “company” and the words in bright orange which are “XYZ” (the company ticker), “strong” and “earnings.” The model places less importance on the association between “company” and words like “today,” and even less importance on words that aren’t highlighted like “as” and “the.”

This differs from other methods of text analysis that are limited to processing information sequentially and tend to overemphasize neighboring words—potentially missing important connections between words that are more distantly separated. For example, at the time that Lehman Brothers went bankrupt in 2008, our model had a very negative score on Lehman. This is obviously a good thing, but it turns out that we were right for the wrong reasons. The model was interpreting “fixed” as a negative word, rather than considering that a lot of the conversation at the time was around “fixed income”—so the more that “fixed income” was mentioned, the more it contributed to a negative score based on just the word “fixed.” This is the problem with analyzing each word in isolation. Because of this, transformer-based models tend to provide a more accurate and precise understanding of the text. Just like ChatGPT can use this technology to predict the next word in a sentence and produce human-like content, we can leverage it to improve our investment insights and predictions.

LLMs are trained using a vast number of data inputs. This is what allows ChatGPT to perform a wide range of tasks and closely simulate human reasoning with broad applicability. By comparison, the LLMs used in our investment process are designed to complete specific investment tasks, such as forecasting the market reaction following corporate earnings calls, for example. As a result, our models are trained on a smaller set of data inputs but are expected to deliver a high level of accuracy in performing the specific task that they’ve been trained and fine-tuned for.

Said another way, because we can marry our expertise and insights within our own models, we are able to create a higher level of accuracy than if we were to use an open-source model less trained on a specific task. For example, Figure 3 illustrates the performance of our earnings call model compared to OpenAI’s larger GPT models at predicting post-earnings market reactions. While the accuracy of OpenAI’s model improved considerably from GPT-3.5 to GPT-4, both OpenAI models demonstrate a lower level of predictive performance than our proprietary model that’s been trained and fine-tuned for this particular purpose.

/blk-inst-c-assets/documents/charts/ai-revolution-in-retirement-2.csv bar-chart column-simple false

Fig 3: Earnings call model fine-tuned to predict at high accuracy

Accuracy at predicting post-earnings market reaction

Source: BlackRock Systematic, as of May 2023. This analysis is based on a sample of 200 earnings calls. The analysis computes the prediction for each model and compares it with return outcomes (positive or negative) based on future 3-day stock returns. The accuracy is computed as the fraction of predictions that were correct for each model.

The punchline here is that the stakeholder you choose to work with matters. We believe our strategic approach is different. As investors, we focus on generating alpha by maintaining an information advantage in markets. We’ve been researching and integrating AI, machine learning, and NLP technologies into our investment process as early as 2007.¹ And our continued commitment to innovation in the fields of finance and AI was further advanced through our development of AI Labs in 2017.

We understand the value of AI and we continue to invest in talent, development of advanced AI models, and the hardware infrastructure to power our proprietary AI models. This is how we can extract unique insights and provide meticulous risk control on behalf of our clients.

2. AI’s potential impact on our ability to save for retirement

It's not hard to imagine a world in which labor productivity can be improved through AI, whether by scaling up medical diagnoses, achieving fully autonomous shipping and transportation, or significantly increasing the productivity of goods manufacturing with more advanced robotics. Suffice to say, this will have monumental implications on which industries are likely to drive long-term returns for retirement savers, but it’s also important that we as retirement practitioners consider the impact on labor markets tied to these industries.

One potential impact is the heightened risk of downside shocks on certain individuals’ abilities to save for retirement should they work in industries or jobs that could be threatened by AI’s ability to automate specific roles. Plan fiduciaries can help their participants augment these possible savings shocks with better return potential from their retirement investments through active management. In addition, integrating insurance into retirement portfolios can help to hedge against human capital risks with better protection of financial capital.

It is reasonable to say that AI’s impact on participants’ abilities to work and save over a traditional timeline (e.g., ages 20-65) may very well change, which could create a potentially wider dispersion of outcomes when we think about retirement readiness. As a leader and innovator in lifecycle research, our Retirement Solutions team, responsible for managing $498bn² in assets across lifecycle solutions, is actively researching how this may affect our approach to lifecycle modeling which ultimately informs the glidepath decisions we make on behalf of millions of retirement savers.

3. AI as a differentiator for your retirement plan

The speed of the AI evolution necessitates dedicated resources and expertise on the part of an asset manager as well as other constituents that—when combined—can result in a robust retirement plan. Yet not all firms are equipped to properly leverage AI, not just as a tool, but as a catalyst for pushing the boundaries of investment insights and other aspects of plan design.

It's important to recognize that integrating AI does not necessarily mean rehauling your entire plan design. In fact, we believe the opportunity to integrate AI into your plan can enhance many of the objectives you are already focused on, which may include enhancing investment returns, expanding participant communications, or improving the efficiency of administrative tasks.

As an example, mundane administrative or oversight tasks like non-discrimination testing have the potential to be automated through machine learning and LLMs.³ This could lead to more time and resources for plan fiduciaries to focus on other aspects on plan design such as the adoption of new safe-harbor regulations or financial wellness programs, which could lead to better financial outcomes for participants.

Other applications include powering participant servicing through sophisticated chatbots that leverage trained LLMs to deliver a fundamental shift in the user experience for participants. Much of this is still in the early stages of development and adoption, but different use cases to enhance the participant experience include identifying relevant financial education programs⁴ or providing personalized financial advice,⁵ or simply providing help to navigate the benefits platform.⁶

Certainly, regulation on the matter of incorporating AI into retirement plans is bound to evolve and plan fiduciaries will need to maintain their expected duty to act prudently when evaluating the appropriateness of any AI application in plan. In the meantime, we encourage all plan fiduciaries to consider the potential of AI and identify areas of application that will be most beneficial to you and your participants.

Conclusion

This article’s primary focus is on the integration of AI as an investment insight within retirement portfolios. But we also see the broader application of AI within the retirement industry through client servicing, participant engagement, and plan design and administration. Regardless of a plan fiduciary’s area of focus, one thing remains clear: Ensure asset managers, recordkeepers, and consultants are equipped to help you deliver more value to your participants through the integration of AI with your plan.