Finance did not need Harry Markowitz to discover diversification. It needed him to turn a familiar investment instinct into a theory of how risks belong together. Language models may now be confronting a similar problem: AI models have become remarkably good at reading, classifying and predicting at the level of individual securities, begging the question on how those outputs should be used in portfolio construction and risk management.
As a graduate student at the University of Chicago, Markowitz read John Burr Williams’s The Theory of Investment Value, which valued stocks by the present value of their expected future dividends. But this raised a puzzle: if expected return were all that mattered, why not put everything into the security promising the most? Sensible investors did not. They diversified.
The observation sounds obvious now because Markowitz helped make it obvious. Investors already analysed securities and diversified, but they lacked a formal way to ask what happened when assets were considered together. A portfolio could not be understood simply as a collection of attractive investments: their risks might reinforce or offset one another. The relationship between them mattered.
That shift became the foundation of modern portfolio theory and shows an intriguing way to think about artificial intelligence in finance today.
When the Analyst’s Language Became Data
Financial researchers have been turning words into quantifiable data for almost two decades. In a 2007 Journal of Finance study, Paul Tetlock quantified pessimism in a Wall Street Journal column and demonstrated its relationship with market prices and trading volume. A year later, Tetlock and colleagues extended the approach to company-specific news, finding that negative language helped forecast earnings while prices initially underreacted to the information it contained.
Yet extracting financial meaning from language was not straightforward. In 2011, Tim Loughran and Bill McDonald showed that almost three-quarters of the words labelled “negative” by a widely used general-purpose dictionary were not normally negative in financial contexts. A “liability” is not necessarily bad news simply because every day English makes it sound ominous. Their finance-specific dictionaries improved the measurement, but the broader lesson was more important: a method could process language consistently while still misunderstanding what that language meant economically.
Meanwhile, computational linguistics was developing new ways to represent language itself. Word2Vec and related models made it possible to compare words with related meanings even when they did not literally overlap. The significance was not that computers had suddenly discovered meaning, but that relationships between concepts could increasingly be quantified. This helped transform search, allowing systems to identify relevant information based on meaning rather than exact keywords.
For finance, these advances opened the possibility of understanding not just which words appeared in a document, but what they referred to and how they related to one another. Aspect-based sentiment analysis, for example, could distinguish optimism about a company's sales from concern about its margins, while contextual language models made representations sensitive to surrounding text. Financial-domain models such as FinBERT brought these developments back to Loughran and McDonald's original insight: understanding language in general is not quite the same as understanding it in finance – context matters.
From Words About Companies to Relationships Between Companies
But understanding what language says about individual companies is only part of the story. The same information can reveal how companies relate to one another.
In 2010, Gerard Hoberg and Gordon Phillips used companies’ own 10-K product descriptions to construct time-varying industries in which each firm could have its own set of competitors. Rather than relying on fixed industry classifications, they allowed a company's peers to emerge from what businesses said they actually did.
Travis Box later showed that similarities in newswire coverage could help predict which companies' returns would move together, even after accounting for conventional measures of economic and financial connection.
Together, these studies suggest a broader possibility: financial text is not merely a source of signals about individual companies, but a way to uncover the relationships and dependencies between them.
Language Models Make the Opportunity Much Larger
Modern large language models extend these possibilities dramatically. A recent Journal of Financial Economics study by Alejandro Lopez-Lira and Yuehua Tang shows that LLM assessments of financial news can capture immediate market reactions and predict part of the subsequent drift. Notably, this advantage declines as LLM adoption rises, consistent with information being incorporated into prices more quickly.
But prediction is only part of the opportunity. If language models can extract financially meaningful information from news, they may also help reveal the economic relationships embedded within it. The question is no longer simply what a model can tell us about an individual company, but what it can tell us about companies considered together.
The challenge, then, is not whether AI belongs in portfolio management. It already does. It is whether the way we use language-derived information has caught up with the sophistication of the models producing it.
From Representation to Portfolio Model
In neuroscience, meaning emerges from patterns across populations of neurons rather than residing in any single cell. Language models work in a similarly distributed way: as a model reads an article, it produces a representation that can be compared with those of other documents. This principle underlies modern semantic search, but it also opens the possibility of comparing the information surrounding entire companies.
In our recent work, we represent each article separately and treat the information surrounding a company as a distribution, rather than reducing it to a single sentiment score or industry label. To compare those distributions, we draw on the work of another Nobel laureate, Leonid Kantorovich, whose mathematics addressed how resources could be moved efficiently between different arrangements. Here the question is similar: how much would one company's information need to shift to resemble another's?
The financial question is whether those relationships help explain which companies move together in markets, and whether they can inform portfolio risk and allocation. The aim is not to replace portfolio theory, but to connect the richer information produced by language models to the financial structures that portfolio theory seeks to understand.
A Markowitz Moment Would Change the Question
Markowitz did not make security analysis, valuation or forecasting obsolete. He changed the unit of analysis: an attractive investment could no longer be considered in isolation. Its relationship to everything else in the portfolio became part of the decision.
Language models may need an equivalent transition.
For an investment committee, it is not enough to know that an AI system reads each company well or predicts each security somewhat better. If ten attractive signals ultimately reflect exposure to the same underlying economic risk, treating them as independent insights may create a concentrated portfolio. Equally, two companies in different industries may be exposed to remarkably similar economic information, a relationship that conventional classifications might overlook.
This becomes particularly important for newly listed or privately held companies with limited market histories but extensive information about their products, suppliers, contracts and competitors. Understanding how that information relates to the rest of a portfolio may be more valuable than another point forecast.
The opportunity for AI in finance is not simply to understand individual companies better, but to understand what connects them.
If you liked this post, don’t forget to subscribe to the Enterprising Investor.
All posts are the opinion of the author. As such, they should not be construed as investment advice, nor do the opinions expressed necessarily reflect the views of CFA Institute or the author’s employer.
Image credit: ©Getty Images