Abstract: Most real-world domains deal with the major changes that may occur after the development of a predictive model. This work shows a novel methodology to detect dataset shift in sets used for logistic regression based predictive modelling. The main difference compared to existing approaches is the use of the estimated coefficients to construct a statistical measure that identifies when predictors are shifting. We provide experimental results using credit scoring datasets, assessing the proposed method's effectiveness and gaining insight of the underlying process that generates the data.
Bio: Cristián Bravo is an Instructor Professor at the University of Talca, Chile, currently on leave as a Visiting Research Fellow at the KU Leuven, Belgium. He is an Industrial Engineer, and has a Master in Operations Research and a PhD in Engineering Systems from the University of Chile. He has served as the Research Director of the Finance Center, U. Chile, as the Chief of Business Intelligence for Consorcio, the largest insurance company in the country, and has been published in several Data Mining and Operations Research journals. His research interests cover Credit Risk, specially applied to Micro-Entrepreneurs, and Data Mining models in this area.