The CEO effect is a long-standing debate in corporate governance. While early research indicates that CEOs significantly impact firm performance, more recent evidence suggests that the CEO effect is negligible in out-of-sample tests. We argue that the CEO effect is not absent but is contextualized and is hidden within textual information that structured data fails to capture. We employ text-embedding techniques to capture CEO biographic information and firm context and train neural networks to identify a contextualized CEO effect. We find that textual information yields a significant 3.54 percentage points increase in out-of-sample R2 for predicting firm performance, even after controlling for standard CEO ability measures. This predictive power is more pronounced during periods of higher macroeconomic uncertainty and in industries with more intensive regulations, indicating the CEO’s important role as a risk manager. Finally, we construct a new CEO ability measure by employing a counterfactual experimental framework based on the match between a CEO and a specific firm context.

zlynne@ceibs.edu