Autocorrelation, structural breaks and the predictive ability of dividend yield

Abstract

We study whether dividend yield (DY) can predict aggregate stock returns while controlling for the effects of structural breaks in the parameters and bias induced by autocorrelation in the predictor variable. To do so we apply the Bai and Perron (BP) (1998, 2000) methodology to test for structural breaks and the bias-adjusted predictability test of Lewellen (2004). We show that although DY predicts market returns during the period 1946 to 1989, there exist ‘natural’ subsamples bounded by statistically detectable structural breaks that can last for long periods of time (up to 11 years in duration) when DY does not show significant forecasting power. This has important implications in that even if in the long-run DY actually provides strong predictive ability, investors should be mentally prepared for long dry spells of unpredictability with respect to DY.

Notes

¹ A sample of some recent examples showing the wide variety of contexts in which structural breaks in economic time series have been analysed include Cook (2005), Sen (2004), Narayan (2005) and Sögner and Stiassny (2002). Cook (2005), for example, examines the unit root test in the presence of structural changes in both the level and variance of integrated time series. Sen (2004) specifies structural breaks to test whether US macroeconomic series difference stationary or trend-break stationary. Narayan (2005) applies unit root structural break tests to investigate whether shocks to Fiji's tourism industry have a permanent effect or a transitory effect on tourist expenditure in Fiji. Sögner and Stiassny (2002) investigate Okun's law for 15 OECD countries and checks for its structural stability.

² See technical appendix.

³ Blanchard and Watson (1982), Tirole (1982, 1985) and Loewenstein and Willard (2000) discuss the theory of rational bubbles; Hamilton and Whiteman (1985), Flood and Hodrick (1986, 1990), Diba and Grossman (1988) and West (1988) provide empirical evidence.

⁴ This is the same method as the one used in Rapach and Wohar (2005).