Model Selection and Paradoxes of Prediction

Abstract:     
In this essay we postulate a number of theoretical hypotheses allowing one to resolve in some degree the following two prediction paradoxes: (1) why simple linear models often have an advantage in predictive power over more complex nonlinear models that lead to a better in-sample fit; (2) why combinations of forecasts often increase the predictive power of individual forecasts. We also give a numerical example illustrating our theoretical statements.

Keywords: model selection, forecasting, linear and nonlinear models, combination of forecasts

JEL Classifications: C10, C22, C53

Working Paper Series