In classical propositional learning, examples are labeled with class values and are described using a set of attributes. The systems learn theories that in some way explain the observed examples of each class. The resulting theories are general descriptions of the classes. These descriptions are logical relations between the attributes' values.
There is an obvious mapping between the variable being predicted in a time series and the class labels of classification problems. They are the goal of learning. We still have the problem that usually classes are a set of finite discrete labels while the values of the predicted variable are a set of possible infinite continuous values. Numeric classes can be a problem to some learning systems. To overcome this difficulty we either use learning systems that cope with numeric classes or we discretize the class values before the learning phase. Learning systems that deal with numeric classes have the advantage of not loosing any information in a process of discretization. The main disadvantage is that few can do that. Some examples are the regression trees like CART [2], M5 [14, 15] and RETIS [8, 12]. In this paper we follow the strategy of using a learning system (M5) that deals with numeric classes.
In order to use inductive learning systems we have to decide which attributes (features) should we use to describe each observed instance of the goal variable. On section 2.2.1 we will see that we can introduce attributes that store relevant information about previous values of the goal variable.