Information disclosed by corporations plays a fundamental role in shaping asset prices and the expectations of investors. That information typically comes in financial statement releases, conference calls, annual reports, and the news media. In a recent paper, though, we examine information from a relatively unexplored source: corporate executive presentations. These presentations are unique in two ways. First, since CEOs must deliver live presentations within a given time limit, their slides tend to include a large amount of visual and graphic elements. Second, complementary to quantitative information and in contrast with other corporate disclosures, executive presentations provide an abundance of visual information about a firm’s product designs and operation plans.

Due in large part to its unique capacity for analyzing alternative data on a massive scale, artificial intelligence is now being deployed with increasing frequency both by the financial services industry and in academia. Motivatedby the potential of AI technologies in processing of visual data from corporate presentations, our paper addresses two central research questions. First, we study whether and how state-of-the-art machine learning techniques can extract valuable information from visual data and help investors understand corporate operations. Second, we examinewhether AI technologies create disparities between market participants by rewarding AI-equipped investors with information and trading advantages.

We first construct a comprehensive data set of corporate presentation slides from multiple sources including Bloomberg News and corporate websites from 2005 to 2018. The sample consists of 17,277 corporate presentation slide decks and 464,765 slide pages associated with an average of 1,023 unique firms per year. Our sample includes multiple types of executive presentation events, including non-deal road shows, IPO road shows, broker-hosted investor conferences, and capital market day events.

To reflect the nature of information conveyed by corporate images, we use a deep learning algorithm to classify them into three categories: 1) Operations Forward are images that provide forward-looking operational information, including future products, blueprints, and development plans; 2) Operations Summary are images that present information about existing products or services; 3) Others are graphs of financial and quantitative information and generic images. In our classification process, we adopt a deep learning model tailored for image recognition: convolutional neural networks (CNN). Furthermore, we use the transfer learning method to improve prediction accuracy and reduce the size of the training sample. Transfer learning uses and fine-tunes existing deep neural network models that have been pre-trained with large, labelled image data sets.

We first examine the value of visual information extracted from corporate presentation images by testing whether visual information can predict stock returns. We hypothesize that Operations Forward contains new, valuable information for investment, whereas the information in Operations Summary will, for the most part, have already been incorporated into stock prices. As expected, we find that Operations Forward is associated with significant and positive short-term cumulative announcement returns (from three days before to three days after the presentation date), while other types of visual information are not. A one-standard-deviation increase in Operations Forward is associated with an increase in abnormal returns of approximately 14 basis points around the presentation date.

Information can affect stock prices via two channels: the discount rate and future cash flows. Given that corporate disclosure of prospective operations is unlikely to be directly related to discount rates, we expect it to influence stock prices through the cash flow channel. Consistent with this hypothesis, we find that Operations Forward is positively associated with firms’ sales and earnings in the fourth quarter and the second year following the presentation. Interestingly, Operations Summary is only significantly correlated with sales and earnings in the nextquarter but not in the long run, indicating that it contains mostly stale and short-term information.

We next explore different market participants’ responses to the visual information contained in corporate presentations. Drawing on recent studies showing that unequal access to alternative data and capacity to analyze data at scale increases information asymmetry among market participants, we hypothesize that the ability to process and extract unique information from unstructured data provides information advantages to institutions that have adopted AI technologies. Since extracting visual information from slides on a large scale requires deep learning capabilities, we expect that AI-equipped financial institutions are more likely to trade on visual information, compared with otherinstitutions and retail investors. Consistent with this hypothesis, we find that institutional investors with high AI investment (as measured by hiring more employees with AI-related skills) trade more around the presentation date when corporate presentations contain more forward-looking visual information related to operations (higher Operations Forward). In contrast, trades by other institutional investors and retail investors are not sensitive to the visual information contained in CEO presentations.

If AI-equipped institutions are better able to process visual information, then stock prices should incorporate such information more quickly when these institutions are present. To test this hypothesis, we separate stocks into twogroups, based on the level of ownership by AI-equipped institutions. We find that Operations Forward is only associatedwith significant abnormal returns around corporate presentations when the stock has a high proportion of AI-equippedinstitutional ownership, suggesting that AI-guided institutional trades are what impound the visual operations information into prices.

Our paper is among the first to provide direct evidence of the value of visual information extracted by AI in corporate disclosures and news media. We show that AI technologies can create information and trading advantages for adopting institutions, potentially generating an AI divide among investors. Moreover, we construct a new data set, comprised of corporate presentations, which, since such information is more abundant and salient therein, provides a unique setting to study visual information about business operations and products. Additionally, our paper explores transfer learning techniques and CNN models, ultimately developing a framework for future studies of visual information in important corporate disclosures.

This post comes to us from Sean Cao at the University of Maryland’s Robert H. Smith School of Business and Yichen Cheng, Meng Wang, Yusen Xia, and Baozhong Yang at the J. Mack Robinson College of Business at Georgia State University. It is based on their recent paper, “Visual Information in the Age of AI: Evidence from Corporate Executive Presentations,” available here.