TernaryrareearthgermaniumsystemswithCuandAg—Areviewandacontributiontotheirassessment.pdfVIP

Predicting Rare Extreme Values Luis Torgo and Rita Ribeiro LIACC-FEP, University of Porto, R. de Ceuta, 118, 6., Porto 4050-190, Portugal {ltorgo, rita}@liacc.up.pt http://www.liacc.up.pt/~[ltorgo,rita] Abstract. Modelling extreme data is very important in several appli- cation domains, like for instance finance, meteorology, ecology, etc.. This paper addresses the problem of predicting extreme values of a continu- ous variable. The main distinguishing feature of our target applications resides on the fact that these values are rare. Any prediction model is obtained by some sort of search process guided by a pre-specified evalua- tion criterion. In this work we argue against the use of standard criteria for evaluating regression models in the context of our target applica- tions. We propose a new predictive performance metric for this class of problems that our experiments show to perform better in distinguishing models that are more accurate at rare extreme values. This new evalu- ation metric could be used as the basis for developing better models in terms of rare extreme values prediction. 1 Introduction In several applications the main focus of interest is a small proportion of the available data. These unusual cases have a large importance, and as such, antici- pating them is a critical task for these domains. An example of such applications is the prediction of the future returns of a stock. Unusually high (low) returns are rare, but they are the most interesting values for investors and thus they should be the target of any financial prediction model. A related problem has been receiving great attention in the