Applying Propositional Learning to Time Series Prediction

In this paper we present some work on trying to apply propositional learning systems to the problem of time series prediction. We describe the difficulties of trying to adapt classification systems to prediction problems. Some techniques to perform this adaptation are presented. These techniques are based on the introduction of new attributes that convey relevant information about the temporal behavior of a time series. These strategies are evaluated on some time series problems. Results indicate that the predictive performance strongly depends on the used set of attributes and that the best set is domain-dependent. Automatic strategies for finding the optimal set of attributes are sketched out. Experiments were carried out in order to compare our performance to the one obtained by other types of methods in the same data sets. The results compare reasonably to other methods, although indicating that more work needs to be done before machine learning-based systems can compete with other techniques on prediction problems.