Abstract
Using machine learning methods, we develop a new measure of aggregate analyst sentiment. We first train analyst-specific neural network (NN) models that capture each analyst's common biases across firms. Using NN model outputs, we decompose the forecast errors of individual analysts into predictable and non-predictable components. Analyst sentiment captures the aggregated non-predictable errors and reflects the "abnormal optimism" of analysts. Using this aggregate sentiment measure, we find that analyst biases vary systematically along the business cycle and are correlated more strongly with household macroeconomic expectations than those of professional forecasters. Further, analysts systematically under-react to macroeconomic information, where they are over-optimistic during recessions and over-pessimistic during recovery periods. A Long-Short trading strategy based on industry-level analyst sentiment earns annualized risk-adjusted return of over 7%