The task of estimating “good” predictive models from available finite (training) data acquired through, e.g., experiments or simulations is common in virtually all fields of science and engineering. The most commonly used predictive models in the case of continuous dependent variable are polynomial regression models acceptable predictive performance of which are usually attempted to achieve by adjusting a maximal degree and using subset selection techniques requiring non-trivial exponential runtime. The paper outlines a different approach not requiring manual degree adjustments and running in a polynomial time and demonstrates its efficiency in a number of applications in the field electrical and control technologies.