| Abstract [eng] |
Mixed Frequency Sampling (MIDAS) regression is based on a restricted lag distribution structure which requires some restrictions, such as smoothness or positivity, to achieve more accurate forecasts. An unrestricted MIDAS regression with Lasso (Least Absolute Shrinkage and Selection Operator), however, aims for better forecasts by simply reducing the dimension of a model. That is why the aim of this paper is to compare forecasting errors of these two regressions, when the restrictions do not represent reality correctly, but the adequacy of model specifications are not rejected. Using the imitation analysis it is shown that while the restricted MIDAS usually gives a better forecast then unrestricted with Lasso, when the real process is unknown and specified inaccurately (but similar enough), the Lasso forecast error becomes smaller then MIDAS while the sample size is increasing. This paper considers problems caused by the restrictions forced on the lag distribution in the new MIDAS context. The results of the research allows a better understanding of the effects of restrictions in forecasting accuracy and offer aid in choosing models for applied work. |
