Bump hunting for risk: a new data mining tool and its applications (Q1861619)

In contrast to usual risk assessment techniques bump hunting is aimed not at the estimation of the posterior risk probability, but to the search of sub-regions where this probability is considerably larger than the average risk over all the explored regions. The authors describe a new forward-backward procedure for searching such critical regions which have rectangular form (boxes). The procedure is applied to a credit data example. A visual tool is described for the control of the procedure. The results are compared to logistic regression, linear discriminant analysis and CART results. The authors also describe a modified version of this procedure which allows censored data analysis. Results of survival data analysis are described.