Can Risks be Good News? Revealing Risk Perception of Real Estate Investors using Machine Learning

Abstract

The SEC mandates firms to inform investors about their assessment of future contingencies in their 10 Ks. However lengthy and complex disclosures – mostly for dozens of firms in an investor’s portfolio – can barely be processed by a human being. To cope with the flood of information, we exploit an unsupervised machine learning algorithm, the Structural Topic Model, to identify the risk factors discussed in 10-Ks. We apply this algorithm to a US REIT sample between 2005 and 2019 to assess whether the probability of appearance of the extracted risk factors helps to explain the perceived risk on the stock market. We find that the majority of risk factors is significantly associated with volatility indicating that our machine-assisted modeling presents a valid approach to quantify risk disclosures in textual form. Furthermore, we investigate in which direction individual topics affect investor risk perception. Even if all kinds of directions exist, uninformative topics with no impact, increasing risk-perception topics, and decreasing risk-perception topics, the latter is clearly predominant. The predominance of the risk-reducing effect indicates that risk disclosures can indeed be considered good news as long as they clarify the implications of already known risk.

Keywords: Real Estate Investment Trust (REIT), risk, text analysis, machine learning, Latent Dirichlet Allocation, Structural Topic Model, 10-K, Item 1A, Item 7A

JEL Classification: C45, C80, G14, G18, K22, K40, M41, M48, R30