High-Frequency Expectations from Asset Prices: A Machine Learning Approach

Abstract

We propose a novel reinforcement learning approach to extract high-frequency aggregate growth expectations from asset prices. While much expectations-based research in macroeconomics and finance relies on low-frequency surveys, the multitude of events that pass between survey dates renders identification of causal effects on expectations difficult. Our method allows us to construct a daily time-series of the cross-sectional mean of a panel of GDP growth forecasts. The high-frequency nature of our series enables clean identification in event studies. In particular, we use our estimated daily growth expectations series to test the “Fed information effect” and find little evidence to support its existence. Extensions of our framework can obtain daily expectations series of any macroeconomic variable for which a low-frequency panel of forecasts is available. In this way, our method provides a sharp empirical tool to advance understanding of how expectations are formed.