The forecasting method with weighted account for observations

This paper investigates how the existence of structural breaks affects the accuracy of forecasts made with a linear regression. Taking into account many observations before the structural break may lead to highly biased forecast, whereas taking into account only post-break observations may lead to a forecast with high variance. In the context of optimal choice between bias and variance the following methods are compared: ordinary least squares applied to all observations and only to post-break observations, the method of optimal selection of the number of observations {introduced by Pesaran and Timmermann (2006, Journal of Econometrics}) to increase the accuracy of the forecast. This work introduces a new method of forecasting where all observations are accounted for with different weights. It is shown that asymptotically the proposed method yields greater forecasting accuracy than other methods. The ease of computation also supports the practical application of the proposed method. Simulations compare the forecasting accuracy of different methods in the presence of structural breaks.