Investment Manager Discussions and Stock Returns: a Word Embedding Approach

Background. It has been debated for a long time whether institutional investment managers have superior ability to pick stocks and to time the market. If so, the next question is whether the investment managers deliver their market insights to investors. As more and more investors delegate their portfolios to investment managers in the U.S. financial market, the questions above are critical to understanding the value created by investment professionals. Aim. This paper investigates whether institutional investment managers are capable in predicting market aggregate returns and whether their public discussions contain valuable market information. Data. The stock return data are from the Center for Research in Security Prices database, and the textual data are letters to shareholders extracted from N-CSR(S) files from the Security and Exchange Commission Electronic Data Gathering, Analysis and Retrieval database. The N-CSR(S) files are annual (semi-annual) certified shareholder reports of registered management investment companies. Methods. I quantify textual documents by mapping words and documents into a low dimensional vector space using the continuous bag-of-words (CBOW) neural network model. Then I use the document vectors to predict value-weighted market portfolio returns using elastic-net. Results. The out-of-sample predictions show that the root mean square error can be reduced by about 6.6% when document vectors are included in the prediction model, in comparison to benchmark models including a constant, a momentum factor and a value factor. The in-sample regressions show that when the proportion of risk aversion related words increases by 1%, the expected annual stock return increases by 1-5%, which is both statistically and economically significant. Conclusions. Investment managers have insights to predict market aggregate returns, and they convey valuable information to their investors in the letters to shareholders in their regulatory reports. The CBOW neural network wordembedding model provides an efficient way to retrieve information from textual documents. Textual features that predict stock returns contain information about the degree of risk aversion of investors.