Can switching between predictive models and the historical average improve bond return predictability?

We propose a novel and simple “switching” approach to improve out-of-sample evidence of return predictability: when the return forecast from a predictive model is negative, we switch to use the return’s historical average as our forecast. When applied to predict Treasury bond returns, this approach can produce stronger evidence of statistical predictability and higher real-time economic gains than the original forecasts. We also show that our approach outperforms the “truncation” approach which replaces negative forecasts with a zero value. Our findings lend support to the hypothesis that predictive evidence exists only in short-lived periods.