Construction of an optimal, of highest precision and stable regression models is considered. A new algorithm is presented which starts by selecting the independent variables included in a linear model. If such a model is found inappropriate, increasingly more complex, higher precision models are considered. These are obtained by addition of nonlinear functions of the independent variables and transformation of the dependent variables. The proposed algorithm is incorporated in the SROV toolbox (Shacham, M. and N. Brauner, 2002, Computers chem. Engng., in press). Using an example, it is demonstrated that the algorithm generates several optimal models of gradually increasing complexity and higher precision from which the user can select the most appropriate model for his needs.