Print
E-MailAnastasia Zakolyukina is an Associate Professor of Accounting and a William Ladany Faculty Scholar at University of Chicago Booth School of Business, Ian D. Gow is a Professor of Accounting at the University of Melbourne, and David F. Larcker is the James Irvin Miller Professor of Accounting, Emeritus, at Stanford Graduate School of Business. This post is based on their recent paper. Related research from the Program on Corporate Governance includes What Matters in Corporate Governance? (discussed on the Forum here) by Lucian Bebchuk, Alma Cohen, and Allen Ferrell.
A significant body of research on corporate governance has emerged in recent decades. Much of this research has focused on individual governance provisions, such as staggered boards or CEO duality. Yet, a careful reading of this research suggests that for most governance provisions, the evidence is mixed. Some papers will find that a provision is good for shareholders, while other papers will find that it is bad. Often later papers attempt to synthesize research and find that the evidence is mixed at best. (See Larcker and Tayan [2020] for discussion of prior research on corporate governance.)
Some papers have looked to incorporate individual governance provisions into broader measures of corporate governance quality. Typically these measures will involve aggregation of governance provisions into a kind of index. But again the evidence is often mixed.
Our paper takes a different approach. We take data on individual governance provisions and a number of firm outcomes and feed these to a machine-learning algorithm with a goal of asking whether measures based on these provisions can predict firm outcomes. We collect comprehensive data on over a hundred of corporate governance features from Equilar, WhaleWisdom, and FactSet. These data cover institutional investor holdings, anti-takeover provisions, the compensation of executive and non-executive board members, the financial expertise of a board, and other board characteristics. We further combine these data with firm characteristics that prior literature identifies as being associated with the outcomes we consider. For firm outcomes, we consider accounting restatements, class-action lawsuits, business failures, operating performance, firm value ], stock returns, and credit ratings. Extant corporate governance research examines all of these outcomes.
We train models that use current characteristics—including over a hundred corporate governance variables—to predict future outcomes. We use gradient boosting of regression trees with cross-validation, an approach that easily accommodates both non-linearities and interactions between variables and can thereby implicitly aggregate provisions into measures optimized for prediction (see Hastie et al. 2000).
Corporate governance researchers rarely test for the out-of-sample predictive value of the causal variables studied. Yet understanding a phenomenon requires both explanation and prediction. Watts [2014] suggests that it is self-evident to many social scientists and philosophers that causal knowledge should facilitate prediction using new data.
Our approach should allow us to detect if any individual governance provisions are useful in predicting firm outcomes, as the algorithm used is flexible enough to discard features that do not have predictive value, but retain those that do. We should also be able to detect if there exist suitably aggregated combinations of governance provisions (i.e., governance indices) are useful in predicting firm outcomes, as the algorithm used is able to combine provisions in a flexible manner.
Unfortunately, for virtually all outcomes, we find that corporate governance characteristics do not help to predict firm outcomes. Using the last three years of our sample as an out-of-sample test period, we find that models based on firm characteristics alone perform as well as models that add corporate governance variables. Moreover, models based on corporate governance characteristics typically fare worse than models that include only other firm characteristics. Our results suggest that there is little or no support for the existence of strong causal relations between corporate governance—as typically measured—and firm outcomes. Given the nature of our algorithms, this conclusion holds for both individual governance provisions (e.g., staggered boards or CEO duality) and comprehensive measures based on aggregating those provisions. In addition, our results suggest that it is very difficult to construct aggregate measures of corporate governance that have predictive value, consistent with Daines et al. [2010], who find that the then-extant commercial corporate governance measures had little or no value for predicting firm outcomes.
