Real-Time Macro Information and Bond Return Predictability: A Weighted Group Deep Learning Approach

Abstract

Relying on a weighted group neural network model, this paper reexamines whether treasury bond returns are predictable when real-time, instead of fully-revised, macro information is used. Two types of real-time macro information are taken into account: macro vintage data and news-based topic attention. For non-overlapping bond returns, the evidence for predictability is weak. When forecasting overlapping bond returns, using only traditional macro variables, the neural network improves over standard machine learning tools for forecasting short-maturity bonds, but this predictability can hardly be translated to economic gains in the presence of realistic leverage constraints. In contrast, when adding textual factors to forecast overlapping returns, using the neural network becomes crucial and improves forecasting performance for all maturities. Adding news to the forecasters also increases economic gains when large leverage is allowed.

Keywords: Deep Learning, Bond Return Predictability, Real-Time Macro Data, News Topic Attentions.

JEL Classification: C45, C53, G11, G12, G17