New methods for measuring cognitive biases and heterogeneous skill in financial analysts and mutual fund managers

Chapter I develops a formal model of analyst earnings forecasts that discriminates between rational behavior and that induced by cognitive biases. In the model, analysts are Bayesians who issue sequential forecasts that combine new information with the information contained in past forecasts. We estimate the model and find strong evidence that analysts are overconfident about the precision of their own information and also subject to cognitive dissonance bias. We examine the influence of the relative amount of private information as a measure of ambiguity on the magnitude of the biases. The variation in overconfidence between the two groups is consistent with the well-established variations documented in the psychological literature. We also demonstrate a relationship between book-to-market ratios and cognitive bias.; Using individual annual earnings forecasts from 25 countries in Asia, North America and Europe, Chapter II estimates the model developed in Chapter I and quantifies the magnitude of cognitive biases across countries. We find very strong evidence that analysts in all 25 countries are overconfident about the precision of their own information, although cross-cultural variations exist. In particular, overconfidence is highest among Asian countries. We also find significant cognitive dissonance in 15 of the 25 countries. In addition, the highest levels of private information precision are associated with countries with strong shareholder protection laws.; Chapter 3 utilizes a Bayesian Hierarchical model to analyze and study the skill level of mutual fund managers. In contrast with several recent papers that apply empirical Bayes techniques on a fund-by-fund basis, the hierarchical model utilized here explicitly recognizes the joint dependence between fund manager alphas. Estimation of the hierarchical model on a set of actual mutual fund returns over the period 1963–1999 produces strong evidence of skill variability, even after controlling for exposure to the four standard factors, expenses and turnover. Bayesian posterior samples are generated using an independence Metropolis chain algorithm. Analysis of this posterior sample allows one to make direct inferences about properties of the entire population of fund managers.