Working Paper: NBER ID: w11841
Authors: Jacob Boudoukh; Matthew Richardson; Robert Whitelaw
Abstract: The prevailing view in finance is that the evidence for long-horizon stock return predictability is significantly stronger than that for short horizons. We show that for persistent regressors, a characteristic of most of the predictive variables used in the literature, the estimators are almost perfectly correlated across horizons under the null hypothesis of no predictability. For example, for the persistence levels of dividend yields, the analytical correlation is 99% between the 1- and 2-year horizon estimators and 94% between the 1- and 5-year horizons, due to the combined effects of overlapping returns and the persistence of the predictive variable. Common sampling error across equations leads to ordinary least squares coefficient estimates and R2s that are roughly proportional to the horizon under the null hypothesis. This is the precise pattern found in the data. The asymptotic theory is corroborated, and the analysis extended by extensive simulation evidence. We perform joint tests across horizons for a variety of explanatory variables, and provide an alternative view of the existing evidence.
Keywords: No keywords provided
JEL Codes: G12; G10; C32
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| persistence of predictive variables (dividend yields) (G35) | correlation of estimators across horizons (C51) |
| sampling error affects long-horizon returns estimates (C83) | sampling error affects short-horizon returns estimates (C83) |
| long-horizon predictability patterns (G17) | attributed to sampling error (C83) |
| joint tests across horizons reveal no significant results (C22) | individual short-horizon regressions may show significance (C20) |