A Multistrategy Approach to the Classification of Phases in Business Cycles Katharina Morik and Stefan R¨ uping Univ. Dortmund, Computer Science Department, LS VIII {morik,rueping}@ls8.informatik.uni-dortmund.de http://www-ai.cs.uni-dortmund.de Abstract. The classification of business cycles is a hard and important problem. Government as well as business decisions rely on the assess- ment of the current business cycle. In this paper, we investigate how economistscanbebettersupportedbyacombinationofmachinelearning techniques. We have successfully applied Inductive Logic Programming (ILP). For establishing time and value intervals different discretization procedures are discussed. The rule sets learned from different experi- ments were analyzed with respect to correlations in order to find a con- cept drift or shift. 1 Introduction The ups and downs of business activities have been observed for a long time It is, however, hard to capture the phenomenon by a clear definition. The National Bureau of Economic Research (NBER) defines business cycles as “recurrent se- quences of altering phases of expansion and contraction in the levels of a large number of economic and financial time series.” This definition points at the multi-variate nature of business cycles. It does not specify many of the modeling decisions to be made. There is still room for a variety of concepts. – What are the indices that form a phase of the cycle? Production, employ- ment, sales, personal income, and transfer payments are valuable indicators for cyclic economic behavior. Are there others that should be included? – What is the appropriate number of phases in a cycle? The number of phases in a cycle varies in the various economic models from two to nine. The NBER model indicates two alternating phases. The transition from one phase to the next is given by the turning points trough and peak. In the model of the Rheinisch-Westf¨ alisches Institut f¨ ur Wirtschaftsforschung (RWI), a cycle consists of a lower turning point, an upswing, an upper turning point, and a downswing. Here, the turning points are phases that cover several months. – Do all cycles follow the same underlying rules or has there been a drift of the rules? There are two tasks investigated by economic theory, the prediction and the dating problem. Where the prediction of values of economic indicators is quite successful handled by macro-economic equations [6], the dating problem remains a challenge. In this paper, we tackle the dating problem: T. Elomaa et al. (Eds.): ECML, LNAI 2430, pp. 307–318, 2002. c  Springer-Verlag Berlin Heidelberg 2002