The ripple effects of the COVID-19 pandemic have increased market volatility and even caused markets to close in some countries. These fluctuations substantially affected mutual funds, leading to fire sales of their assets and SEC scrutiny of their risk management. Investors responded quickly and withdrew more than $40 billion from mutual funds in the first two months of the pandemic. With nearly half of the households in the United States having their pension plans and life savings invested in mutual funds, understanding the risk-taking behavior of mutual funds is thus of prime importance for investors.
For fund managers, the ability to correctly evaluate and manage risk is directly relevant to their goal of achieving high returns. Regulators are also interested in fund managers’ risk-taking behavior to prevent excessive risk, which may exacerbate market volatility during crises. Despite the importance of this subject, little is known about fund managers’ forward-looking assessment of risk and how they act upon their opinions.
In a recent paper, we investigate mutual fund managers’ risk sentiment and their subsequent actions. To overcome the challenge of identifying how mutual fund managers perceive their future risk management and investment plans, we use deep learning to extract information relevant to risk management from mutual fund managers’ portfolio discussion in their mandatory SEC filings (shareholder reports in N-CSR/N-CSRS filings). Such information allows us to construct text-based risk sentiment measures of mutual fund managers. These measures serve as a direct proxy for a mutual fund manager’s forward-looking assessment of risk and allow us to answer several questions: Are mutual fund managers’ subsequent risk-taking behavior consistent with their risk sentiment? Can forward-looking risk assessment lead to superior performance? What are the incentives for managers to disclose their risk sentiment?
Conventional risk-taking measures are calculated based on historical numerical data of returns or holdings. While numerical data are easy to process, they do not contain forward-looking information of managers’ own risk sentiment. In contrast, while managers’ qualitative discussions of portfolio decisions in SEC filings convey rich information, there are challenges in processing them. The textual data of discussions about portfolio decisions in shareholder reports are unstructured and qualitative. Managers can discuss their risk assessment in a variety of linguistic styles, making it difficult to identify patterns. Consider a hypothetical discussion: “Risk is beneficial, but our flow is constrained in the next period.” In this case, the traditional“bag-of-words” approach counts words indicating sentiment in the text. Both “beneficial” and “constrained” are captured. Hence, a rule-based risk measure will generate a neutral risk assessment (one positive word and one negative word). However, the true interpretation of the sentence is that it expresses a positive risk assessment given that “constrained” is used to modify “flow” instead of “risk.” To this extent, we apply deep learning models to incorporate the syntactic relations among words to capture positive and negative risk sentiment.
We employ deep neural networks for natural language processing developed by Chen and Manning (2014) to parse all textual contents in mutual fund shareholder reports and construct directional risk sentiment measures. We first validate these measures by exploring their capability of predicting future risk-taking after controlling for current risk-taking. We find negative (positive) risk sentiment strongly predicts managers’ reduction (increase) of their risk-taking in the subsequent period. We also compare our deep-learning-based measures with the traditional bag-of-word measures that simply count risk-related keywords. Our measures yield results superior to the bag-of-words measures in predicting the future risk-taking behavior of fund managers.
We further find that fund managers who are conscious about negative risk, i.e., having a negative risk sentiment, are able to generate higher future fund performance, measured by Fama-French Carhart four-factor alpha (Fama and French, 1993; Carhart, 1997) and Sharpe ratio (Sharpe, 1994), suggesting that our negative risk sentiment measure extracts valuable information from managers’ disclosure. Since the measure is forward-looking, it can help fund investors to ex ante identify managers’ risk management plans and predict future returns.
Skilled managers may be able to identify better information about risk and thus are more likely to take their own advice. Following Kacperczyk, Sialm, and Zheng (2008), we use the return gap of mutual funds as a proxy for managerial skill. We find that negative risk sentiment helps select skilled managers in the future. Meanwhile, the reduction in risk-taking is more pronounced when a manager is skilled, and only a skilled manager is capable of translating her risk sentiment into superior performance.
The textual discussions in the shareholder reports do not follow any templates, giving managers a large degree of freedom on what to convey to investors. We next explore the incentives of managerial disclosure of qualitative information to investors. We find risk-conscious managers who report negative risk sentiment receive higher Morningstar ratings in the future. The increased Morningstar ratings are largely driven by the managers’ future reduction in risk-taking. Furthermore, funds with sophisticated investors are reawarded with more capital from investors when disclosing negative risk sentiment.
While we focus on situations where fund managers are required to disclose their investment decisions, there may still be discretion in the extent of fund managers’ disclosure. The disclosure literature suggests that managers may choose to disclose risk-related information for several reasons. First, hiding information may result in an investigation by SEC and trigger lawsuits. Second, disclosing risk sentiment information is less likely to attract copycatting behavior than other types of disclosure, such as corporate strategies or portfolio holdings (Cao, Du, Yang, and Zhang, 2021). Third, disclosing valuable information that proves accurate in the future signals managerial skill and can help the manager in the future labor market.
Our paper is the first to develop measures of forward-looking risk assessment of professional money managers, which can inform investors and researchers about fund managers’ risk management and investment decisions. In constructing our text-based measures, we also note that transfer learning, or building special-purpose models based on pre-trained general deep learning models, can save time and solve the challenge of limited training data for machine learning models. Overall, we believe it is promising to develop more applications of deep learning models in textual analytics that can reveal and analyze linguistic features previously inaccessible to researchers.
Cao, Sean S., Kai Du, Baozhong Yang, and Alan L. Zhang, 2021, Copycat skills and disclosure costs: Evidence from peer companies’ digital footprints, Journal of Accounting Research 59: 1261-1302.
Carhart, Mark M, 1997, On persistence in mutual fund performance, Journal of Finance 52, 57-82.
Chen, Danqi, and Christopher Manning, 2014, A fast and accurate Dependency Parser using Neural Networks, Proceedings of EMNLP.
Fama, Eugene F., and Kenneth R. French, 1993, Common risk factors in the returns on stocks and bonds, Journal of Financial Economics 33, 3-56.
Kacperczyk, Marcin, Clemens Sialm, and Lu Zheng, 2006, Unobserved actions of mutual funds, Review of Financial Studies 21, 2379-2416.
Sharpe, William F, 1994, The Sharpe Ratio, Journal of Portfolio Management 21, 49-58.
This post comes to us from professors Sean Cao and Baozhong Yang at Georgia State University’s J. Mack Robinson College of Business and Alan L. Zhang at Florida International University’s College of Business. It is based on their recent article, “Deep Learning Mutual Fund Disclosure: Risk Sentiment, Risk Taking, and Performance,” available here.