Copilot, not autopilot: How generative AI augments, but doesn't replace active management

Artificial intelligence (AI) is commonly defined as machines that mimic the cognitive functions of the human brain. For some cases, like playing checkers, where the rulebook is simple, this is a relatively low bar. Indeed, this was one of the first use cases of AI by Arthur Lee Samuel in 1952.¹ However, the bar rises exponentially with every notch of complexity. It wasn’t until 1997 that Gary Kasparov would lose a full match to Deep Blue.² It took nearly two more decades, even with all the exponential strides in computing power over that time, before AlphaGo beat Go grandmaster Lee Sedol in 2016.³ Thus, while advances in AI, including the ones we’ll discuss in this article, are expanding Earth’s collective cognitive ability, it is premature to seek shelter from sentient robot overlords or even fear that they’ll fully replace many knowledge workers, such as investment professionals.

Instead, with the advent of large language models (LLMs), which are deep learning algorithms trained on gigantic datasets, AI’s output can range from concise summaries to detailed insights. What may first come to mind is OpenAI's GPT-3, of which ChatGPT is the result.⁴ GPT-3 was trained on nearly the entirety of the internet and most books.⁵ This gave its neural network 175 billion parameters,⁶ which it uses to opine on topics ranging from banal to sublime. With terabytes of training data, extensive power amplified by distributed computing, and some old-school human ingenuity, the applications of AI to many fields, including investing, will continue to rapidly advance. While many of these are beyond the scope of this introductory article, we present use cases of how AI can be harnessed by different investors to potentially improve their desired outcomes and workflows.

AI capabilities: Data analysis and predictive power

Distilling investing to an extreme, we could say it is determining the fair value of assets—based on analyzing as much public information as can be gathered—and then, if prevailing market prices differ from the results, buying or selling them. The sheer amount of relevant data is vast—financial documents, earnings transcripts, regulatory filings, news articles, day-long congressional testimonies, and nowadays even Reddit conversations and tweets. This data is noisy, non-normal and increasingly unstructured (that is, inherently difficult to analyze). LLMs can both consume and, critically, understand this data at rates eclipsing any analyst team.

A basic output of this task is the ability to summarize information for human consumption—whether it’s thousands of social media threads written in zoomer vernacular (no cap⁷) or the dense legalese of a corporate deposition (veritably⁸). Taking it a step further, AI can combine different data sets to extract insights not immediately apparent to even a seasoned human investor.

So, should we all retire and let the machines take over? Not so fast. When properly prompted LLMs are quick to offer answers with the confidence of an economist spouting talking points on TV. This is because LLMs are trained, on a Pavlovian level, to offer responses humans will trust. There is a reward function in most algorithms for providing acceptable answers. But is their confidence justified? This depends on many factors, and even if fed high quality data, deep learning algorithms are fallible. For example, transformer models (which construe most LLMs) can easily veer off track, or hallucinate, because they work by sequentially predicting the next most probable word in a sentence. This is an autoregressive process, where words the LLM generated itself are used to predict the next ones. While at first it sounds similar to how humans think—after all the words we say next are predicated on the ones that just left our mouths—LLMs have a much harder time realizing if they are talking nonsense. Recognizing when confident-sounding AI is abjectly wrong, phrasing questions for it with precision, fine-tuning its training, and feeding it the most nutritional data are all reasons for why humans remain a crucial part of the process. We offer practical examples in the world of investing below.

Use cases for asset management, wealth management, traders and retail investors

Investors of all stripes can potentially benefit from using AI. The technology will not put retail traders on equal footing with institutional investors because they do not have access to the disproportionately expensive and often proprietary data on which to train an AI system that institutional investors have been cultivating for decades; nor would they typically know how to fine-tune deep learning algorithms to maximize their potential. Still AI can boost most investors’ scale, speed and sophistication.

Asset management use cases:

Portfolio managers can train LLMs on earnings calls, stock price movements, news articles and social media chatter. They can further input information on behavioral biases (the theory that inefficiencies in the market exist due to human irrationality), their own research notes, security ratings and trade executions. After training this data can be piped into the LLM in real-time. That, in turn, leads to several novel applications such as:

• Combining the sentiment expressed via unstructured information (tweets, subreddits, analyst reports, news, etc.) with structured data (company fundamentals, consensus forecasts, macro indicators) to identify incongruencies that may lead to large price moves.
• AI can help risk managers by providing early warnings of market shocks inferred from secondary and tertiary effects. For example, imagine a fixed income portfolio where some positions’ credit spreads begin rapidly widening. A human manager would immediately understand the increased risk to those underlying positions but what about the rest of the portfolio? An AI with billions of optimized synapses could predict which issuers could be the next domino based on a multitude of data points—from time series correlations to news articles to 10-Ks (company annual reports). A recent, though tragic, example would be the invasion of Ukraine leading to a sudden contraction in neon gas exports, a key component of automotive semiconductors affecting chipmakers, which then impacted carmakers. A well-trained neural network could find this complex linkage the moment the first mortar hit Mariupol, something few humans did.
• AI can alert portfolio managers if their desired trades exhibit behavioral biases. For instance, according to the disposition effect, some investors are reluctant to sell losing positions, yet happy to shed assets that just had big price pops. Differentiating between a prudent decision supported by valuations and one driven by emotions, like regret avoidance, is where an AI trained on previous trades and behavioral finance can take the role of an unbiased coach.
• Because LLMs can process conversational queries, the knowledge moats for doing complex investment tasks—like multiperiod optimization, strategy simulation and factor decomposition—are drying up. In a way, generative AI is democratizing some of the superpowers quantitative (or quant) investors previously hoarded. Soon a multi-asset portfolio manager could ask their AI copilot to “construct a portfolio that is most resilient to a US Federal Reserve pivot, but could still offer 4% yield, is not overweight the growth factor, and wouldn’t have had annualized risk greater than 17% over the last five years” and get a model back. Provided, of course, that one could be constructed with those hurdles. While we know there are no guarantees to achieving these outcomes, we are working on building such a tool at Franklin Templeton Investment Solutions.

Sustainable investing use cases:

Environmental, social, and corporate governance (ESG) analysts could train their AI system on the sustainability disclosures of public companies, quantifiable ESG metrics and press releases about a company’s ESG declarations.

• The AI could then attempt deciphering whether popular beliefs about a company’s ESG practices match reality, or whether companies practice what they preach on any number of sustainability metrics, such as equal pay, carbon footprint reduction and board independence.
• By analyzing data that has not yet flowed into disclosures, AI can identify which companies are making improvements in their ESG practices. Identifying such ESG improvers early may yield better investment results. For example, what if a company plagued by controversy over its treatment of minority employees started putting in diversity, equity and inclusion (DE&I) language into its newest job postings? AI can generate inferences from data points like this nearly in real-time when properly tuned and trained.

Wealth management use cases:

Financial advisors can use AI systems to maximize their clients’ ability to meet the goals most important to them—a task that often involves more than just maximizing return for a given level of risk.

• Determining client investment objectives and risk tolerance is often done through straightforward questions. But how well do clients know themselves, especially under duress? Advisors often receive panicked phone calls from clients after small market drops, demanding “corrective” action, even from those who say they can tolerate large market swings. These requests are often against the clients’ interests. AI trained on past interactions can look beyond surveys and better predict client behavior to suggest portfolios most likely to keep them invested through volatility and even recommend opportunities to proactively reach out, before panic sets in.
• Just as self-reported risk tolerance may not match reality, client financial goals are often not appropriately prioritized. AI can analyze client consumption patterns, needs and desires to chart a dynamic path most likely to maximize the chance of achieving their highest priority goals, while minimizing the chance of running out of money. This is an area where we have been pioneering AI use since 2020, by creating a solution that makes personalized asset allocation and consumption recommendations.

Retail investor use cases:

Most examples above require extensive proprietary data and the knowledge to train and tune models. It bears repeating that while AI is a step toward democratization of investing, it is not an equalizer. Without terabytes of quality data, real-time feeds and superlative computing power even sophisticated retail investors will be at a disadvantage compared to institutions. Nonetheless commercially available AI models can still benefit them.

• Even more so than institutional portfolio managers who generally have risk managers looking over their shoulder, AI may alert retail investors to behavioral biases they may be exhibiting based on the context surrounding their trading. For example, are they entering an option position where risk could far exceed the equity trades they’re used to?
• AI may help create insightful charts, with topical overlays to give visual context to earnings announcements, economic regimes, sector profit margins and possible payouts for a trading strategy.
• LLMs can extract key concepts from lengthy documents—like management commentary or central banker speeches—to help retail investors grasp key concepts.

Conclusion

Understandably, there is both fear and excitement around the advent of AI, and as with most breakthroughs, the nuanced truth should evoke some of both. While AI may create negative externalities, eradicating humankind is unlikely to become its agenda; and though it will augment our lives, it won’t create utopia. For now, in the world of investing, it can fill the role of a tireless junior analyst or unbiased coach, as illustrated by the case studies above. By partnering with algorithms, investors may produce better returns, mitigate risk, reduce their irrational impulses and may come closer to achieving their financial goals.