A Consensus-Bottleneck Asset Pricing Model

Abstract

In this thesis, I introduce a Consensus Bottleneck Asset Pricing Model (CB- APM), a novel empirical framework that employs a partially interpretable neural network to understand and imitate the reasoning processes of sell-side analysts. By utilizing advanced machine learning techniques, CB-APM leverages firm-level and macroeconomic information to predict future risk premiums of U.S. indi- vidual equities. The model demonstrates superior performance in long-horizon predictions compared to a naive neural network-based return prediction models, notably enhancing both predictive accuracy and interpretability. Furthermore, the model’s portfolio performance significantly surpasses that of a buy-and-hold strategy, maintaining robust profitability even during recession periods. CB- APM provides detailed insights into the impact of analysts’ consensus on asset pricing, contributing a novel perspective to the field of financial economics.