Searching for the DGP when forecasting : is it always meaningful for small samples?

Abstract

In this paper the problem of choosing a univariate forecasting model for small samples is investigated. It is shown that, a model with few parameters, frequently, is better than a model which coincides with the data generating process (DGP) (with estimated parameter values). The exponential smoothing algorithms are, once more, shown to perform remarkably well for some types of data generating processes, in particular for short-term forecasts. All this is shown by means of Monte Carlo simulations and a time series of realized volatility from the CAC40 index. The results speaks in favour of a negative answer to the question posed in the title of this paper.