Abstract:
We compare the out-of-sample forecast accuracy of the long-horizon regression
and the first-horizon regression when the regressor is correlated with the regression error.
We show that the long-horizon forecast is more accurate than the first-horizon forecast in
terms of mean square forecast error when the time-series dimension data goes to infinity.
The first chapter, following the introduction, demonstrates this result in the univariate timeseries
context. The following chapter extends the analysis to the panel data context.
Theoretical derivations are supported by an empirical application in chapter four. Given that
regressor-error is a testable statistical property of the regression model, our results give the
practitioner a priori conditions by which he can select the regression model that will yield
the most accurate out-of-sample forecast.