ROP Model: Numerical and Nonparametric Classification-regression

Variability is the fundamental concept of the statistical approach. It is expressed most often by quantitative parameters of which the best known is the variance, quadratic deviation from an average. We propose to express this variability in relation to an average effect, by directly identifying the observations that are responsible for it. For this we have developed an innovative regression-classification model in which each node is managed not by a single preacher but by a risk score including all preachers. The final classification tree is a combination of a network and a tree. The analysis of the variation of the coefficients of the risk scores makes it possible to identify groups of observations for which a preacher completely changes the direction of his prediction. We make the hypothesis that these observations have a particular information missing from the analyzed data, which we call phantom factor, that the others do not have and which would explain the radical change of the meaning of the preacher. Identifying this phantom factor amounts to modeling information gaps, the result of which is the updating of new information unknown to the analyzed data. The computation of the risk coefficients for each predicator at each node is performed by a numerical combinatorial analysis which requires computational power inaccessible before the era of the teraflopic computing stations. The ROP model opens up many perspectives that are directly proportional to the computational power. Several applications of the use of this model have been published in indexed journals.