Skip to main content

Faculty & Research


Bayesian Herd Detection for Dynamic Data

Journal Article
This article analyzes multiple agents who forecast an underlying dynamic state based on streams of (partially overlapping) information. Each agent minimizes a convex combination of their forecasting error and deviation from the other agents’ forecasts. As a result, the agents exhibit herding behavior, a bias well-recognized in the economics and psychology literature. The authors' first contribution is a general framework for analyzing agents’ forecasts under different levels of herding. The underlying state dynamics can be non-linear with seasonality, trends, shocks, or other time series components. The authors' second contribution describes how models within the authors' framework can be estimated from data. They apply their estimation procedure to surveys of equity price forecasts and find that the agents concentrate 37% of their efforts on making similar forecasts on average. However, there is substantial variation in the level of herding over time; even though herding fell substantially during the 2007–2008 financial crisis, it rose after the crisis.

Associate Professor of Technology and Operations Management