Parameter Learning, Sequential Model Selection and Bond Return Predictability

The paper finds both statistically and economically significant out-of-sample evidence of bond return predictability for a Bayesian investor who learns about parameters, hidden states, and predictive models over time. We... [ view full abstract ]

The paper finds both statistically and economically significant out-of-sample evidence of bond return predictability for a Bayesian investor who learns about parameters, hidden states, and predictive models over time. We find that the factor extracted from a large panel of macroeconomic variables contains rich information on future excess bond returns and that introducing stochastic volatility can improve predictive performance. Interestingly, economic evidence is much more pronounced when we do not impose any investment weight constraints, and there seems to be a squeeze of intermediaries capital and a scarcity of arbitrage capital when investors require extreme long positions. We also document that model combinations work well in predicting excess bond returns