Predicting Rare Extreme Values - recent developments

Predicting rare extreme values of a continuous variable is of key importance in several important real world applications (e.g. finance, ecology, etc.). In this seminar we start by presenting the problem and its motivation and then go through a series of existing approaches to the problem, highlighting their main limitations. As a result of this analysis we then describe a series of recent developments of our work in this area that has lead to the introduction of the notions of cost and benefit surfaces. We present these two new notions and their intuition in the context of handling regression problems with differentiated importance of observations (as it is the case of predicting rare extreme values). We formalize both notions and explain how they can be used in the context of our target applications. We finish by providing some initial results on the use of these notions in the context of evaluating models in tasks related to the prediction of rare extreme values of a continuous variable.