Predictive power: Forecasting dividend cuts with AI

Vanguard’s Quantitative Equity Group, which co-manages one of the largest active equity income funds (Vanguard Equity Income Fund), employs a fundamentally based, data-driven systematic process for stock selection in the high-dividend-yield universe (HDYU). Recently, this group embarked on an experiment to see whether AI could build on our Vanguard Equity Income Model, which informs investment decisions.

At its core, the Equity Income Model seeks to invest in stocks with sustainable and growing dividends, high-quality financial health, positive market sentiment, and attractive valuations. The model exhibits a strong relationship with future dividends and has been effective at helping avoid companies likely to reduce their dividends and identifying companies that are likely to sustain and grow dividends in the future.

As the chart below shows, when we divide the universe of stocks into quintiles based on the model’s scores, with Quintile 1 representing the most attractive stocks and Quintile 5 the least attractive, we observe that there are approximately 2.5 times more “cuts” among companies in Quintile 5 than in Quintile 1. Additionally, firms in Quintile 1 exhibit a median dividend growth rate of 8.4%, compared with 1.2% for firms in Quintile 5.

In the world of investing, even a tiny advantage can affect returns. While companies in the HDYU rarely lower dividends, their stock price tends to suffer when they do, making it advantageous to avoid these companies in strategies focused on dividends.

“Vanguard’s quantitative Equity Income Model is great at capturing numerical information and has shown strong predictive power for future dividend growth,” said Yu Tang, assistant portfolio manager on the Global & Income team. “We wanted to know whether earnings calls could offer additional insights about the strength of a company’s dividend outlook. Recently developed large language models are ideally suited to capture nuanced information that our model may not detect.”

The team used Natural Language Processing techniques to cull relevant dividend discussion snippets from more than 22,000 earnings calls from April 2012 through April 2023 for stocks in the HDYU.

Afterward, team members deployed an open-source large language model (LLM) and prompted the model to analyze the extracted snippets to determine whether companies were likely to grow dividends (a positive outlook), were likely to trim dividends (a negative outlook), or had an unclear outlook on their future dividend.

To understand how this works, let’s consider how the LLM analyzed an actual earnings call. In 2015, an analyst asked a company’s CEO whether its earnings before interest, taxes, depreciation, and amortization were likely to cover the dividend. The executive responded that it was reasonable to argue that the dividend money might be needed elsewhere but didn’t directly answer the question. The language model interpreted this as increasing the odds that the company would decrease its dividend, and just a few months later the company did exactly that. 

The experiment’s results also validated the effectiveness of Vanguard’s quantitative Equity Income Model. In the next chart, companies in the HDYU are ranked into quintiles based on their sustainable dividend growth score in Vanguard’s Equity Income Model. This score is one of the key determinants of the overall model’s opinion on a stock. Quintile 1 represents stocks with the highest ability to sustain and grow their dividends, and Quintile 5 represents those with the lowest.

We observed that the language model was remarkably consistent with the sustainable dividend growth scores from the Equity Income Model: 41% of companies with an LLM-based negative dividend outlook fell into the bottom quintile of the sustainable dividend growth score. Further statistical analysis showed that the language model provided additional insights—companies that the LLM flagged as having a “negative outlook” were almost five times more likely to cut their dividend within the following month than companies not flagged.

This experiment demonstrated that LLMs have the potential to add value in our investment decision process with insights not already captured by the numbers. It’s important to keep in mind that while AI offers many benefits, it also has commonly known risks and limitations for fund management, including AI model risk, overfitting (when the model learns the training data too well and includes random noise), and less transparency than traditional models. (These risks and limitations can be managed in many circumstances.) More work needs to be done for a signal to be made into a factor in our model, but this first step provided the encouragement to continue to exploit this technology in pursuit of benefits for investors.

“Large language models are increasingly blurring the lines between systematic and discretionary investing approaches, making insights that were once confined to meticulous fundamental analysis more accessible and scalable,” said Gil Garcia, an active equity investment analyst on the team.

We continue to identify and evaluate new and exciting technology opportunities like this one to advance what we do and how we serve clients around the globe.