This paper proposes a weighted group neural network model and reexamines whether treasury bond returns are predictable when real-time, instead of fully-revised, macro information is used. Real-time macro vintage data and news-based topic attention are taken into account. We find that news contains rich information on future bond returns beyond traditional macro variables. Our proposed model can help find significant statistical evidence for forecasting non-overlapping short-term bond returns and all overlapping bond returns. Furthermore, the statistical evidence of overlapping bond return predictability can be translated into investors’ economic gains for long-term bonds when investors are allowed to leverage their investments.

To find more, download the full paper.


City University of Hong Kong
City University of Hong Kong
ESSEC Business School, Paris-Singapore
Fudan University