ABSTRACT

This chapter presents a couple of different models in order to make the following discussion more concrete. It shows how the cross validation technique is fairly easily modified into forward validation of time series forecasts. The chapter provides general methods for model selection and comparison of model classes based on such validation and gives some examples. It demonstrates some different loss functions for various kinds of forecasts, the reason being that the imagination should not be limited to just scalar forecasts and squared errors. There are some good reasons for fitting models. One is that forecasts and other analyses become objective and unaffected by optimistic or pessimistic moods. A second reason can be that decisions are made so often that a subjective technique becomes cumbersome and/or will be less carefully made. A third and more important reason can be that the models give better forecasts and will find relations which are not easily observed by eye.