As artificial intelligence (AI) continues to make its way into financial markets and impact corporate decisions, quarterly and annual reports are increasingly featuring machine-friendly words and tone, research from FIU Business shows.
The researchers examined actual speeches by corporate executives during earnings conference calls and found the speakers sound more positive and excited in their vocal emotion if their firm's investors are equipped with AI technologies.
“There’s a change in tone and sentiment,” said Alan Zhang, assistant professor of finance at FIU Business. “The person [who writes the document] knows which words to use, and the machine [the AI reader program] doesn’t classify them as negative.”
Forthcoming in the Review of Financial Studies, it’s the first study to identify and analyze the feedback effect: how company executives adjust the way they write knowing that AI programs are listening.
The findings highlight the growing roles of AI in the financial markets and their potential impact on corporate decisions, Zhang explained.
“Managers want to portray their business activities and prospects in a positive light to attract or gain stakeholders,” he added. “If the firm discloses a negative report, investors respond negatively, and they get returns that are negative.”
If the sound is more excited, it presents a better image to the machine readers.
“As a result, in some cases when the SEC posts the disclosure, trades happen more quickly,” said Zhang. “Investors and professional traders are ready to cut the trade as soon as possible.”
He explained that as public company executives began preparing their SEC disclosures to be machine-friendly since investors began to use AI programs to read those documents. If someone doesn’t want to go through a company’s financial report page-by-page, but would rather listen to a summarized version, the AI program reads it from the text.
“The AI reader is still trained by humans and mimics how a human would analyze and understand a document,” he said. “But it can read filings of thousands of companies more efficiently than a human.”
The researchers examined 359,819 filings from 13,763 companies between 2003 and 2016, recorded by the Securities and Exchange Commission’s Electronic Data Gathering, Analysis and Retrieval (EDGAR) system.
If an IP address downloads filings from more than 50 companies on SEC website within a day, the IP address is classified as machine-reader. SEC publishes the web traffic data for filings on EDGAR, which allows researchers to infer who are machine readers vs. human readers, Zhang explained. So, to conduct their analysis the researchers tracked the IP addresses that were conducting disclosure downloads in large batches.
“The sheer volume of machine-downloaded documents made it unlikely for them to be processed by human readers alone,” said Zhang.
Zhang conducted the research with Sean Cao of the University of Maryland, Wei Jiang of Emory University and Baozhong Yang of Georgia State University.