An investment strategy based on findings culled from a new dataset is proving that it can provide substantially better financial returns than a traditional “buy and hold” approach. 

The dataset compiles meeting minutes, speeches, and press conference transcripts from the Federal Open Market Committee (FOMC). It is the largest tokenized and annotated dataset of its kind.

An investment strategy developed using the dataset predicted investment returns yielding 163.4% higher than the buy and hold method on the QQQ index fund from 2011 to 2022.

The dataset and strategy are part of new research findings from Georgia Tech. The findings document the influence the FOMC has on markets and the economy through its public communications. The research is being presented this month at the 61st Annual Meeting of the Association for Computational Linguistics (ACL 2023).

“By understanding the impact of FOMC communications on market movements, investors can make more informed decisions, and potentially protect their portfolios from sudden downturns or capitalize on growth opportunities,” said Ph.D. student and lead researcher Agam Shah.

“Additionally, it can help economists at the Federal Reserve Banks more efficiently understand the impact of their communication.”

[MICROSITE: Georgia Tech at ACL 2023] 

The dataset contains 214 meeting minutes, 1,026 speeches, and transcripts from 63 press conferences. The meeting minutes and speeches span from January 1996 to October 2022. The press conference archive dates from April 2011 to October 2022.

To explore this heap of FOMC pronouncements, Shah and his team crafted a novel machine-learning classification task. The new task categorized statements in the dataset as hawkish, dovish, or neutral, rather than just positive, negative, or neutral.

The classification task allows computer models to understand FOMC policy stances through the language used in their correspondence. This in turn guides models to predict how markets react to communications, giving investors valuable information to form their own strategies. 

“One of the reasons our research achieved these remarkable results is because it harnesses the power of natural language processing (NLP) to systematically analyze a vast amount of data which is impractical for humans to process effectively,” Shah said. “This provides a much more nuanced understanding of the market’s response to FOMC communications.”

Shah is a Ph.D. student in the School of Computational Science and Engineering (CSE). He is advised by Sudheer Chava, a professor in the Scheller College of Business. Suvan Paturi, a Georgia Tech alumnus and software engineer at Nasdaq eVestment, co-authored the paper with Shah and Chava.

The group will present their paper at a time when the FOMC and the Federal Reserve are in news headlines now more than ever. To curb inflation, the Fed has increased interest rates ten consecutive times from March 2022 to June 2023.

One example that inspired the group occurred during this period on Aug. 26, 2022. Here, FOMC Chair Jerome Powell gave an eight-minute speech that resulted in an almost $3 trillion decline in U.S. equity market value.

This study not only affirms that markets are reactive to words spoken through public communications but now those effects can be measured and predicted. It also provides new tools to help investors make better, more informed decisions. 

“The application of computational methods to finance and economics revolutionizes the way analysts interpret data. It enables us to handle enormous datasets and extract valuable insights that would otherwise remain hidden,” Shah said. 

“This empowers decision-makers to craft strategies that are based on a deeper understanding of market dynamics, leading to potentially higher returns and more efficient financial systems.”