The Predictive Power of the Yield Spread under the Veil of Time

I apply a multiresolution decomposition to the term spread and real-GDP growth in the U.S. Using the filtered data, I study whether the yield spread helps forecasting output. The results show that the predictive power of the yield spread varies largely across time scales both in-sample and out-of-sample at various forecast horizons. Contrarily to the existing literature, I find evidence of a strikingly negative long-run relationship between the spread and future GDP growth over a frequency that spans from 8 to 16 years per cycle. A linear combination among filtered yield spreads shows a sizable improvement in forecasting out-of-sample. The decomposed series are also used for proposing a solution to the breakdown in the in-sample predictive relationship documented by Dotsey (1998) that occurs after 1985.