Title: Long-run predictability tests are even worse than you thought
Authors: Tamas Kiss - Orebro University, School of Business (Sweden) [presenting]
Erik Hjalmarsson - University of Gothenburg (Sweden)
Abstract: The interaction between the two problems of endogenous predictors and inference in long-horizon regressions is studied. The key finding is that long-horizon predictive regressions exacerbates the endogenous predictor bias. Specifically, while endogenous predictors are usually considered problematic only if they are sufficiently persistent, we show that in long-horizon regressions, a version of the Stambaugh bias is present - and substantial - regardless of the (lack of) persistence in the predictor. We derive asymptotic results for a scaled version of the OLS t-statistic. With exogenous regressors, the scaling correctly controls for the overlap in the data, and the scaled t-statistic is very close to normally distributed also in finite samples. With endogenous regressors, the distribution of the scaled t-statistic differs substantially from standard normal. This holds regardless of the persistence in the predictor, and the (asymptotic) Stambaugh bias arising in the scaled t-statistic is thus independent of the persistence in the predictor, and completely induced by the formulation of the long-horizon regression.