IS MULTIPLE LINEAR REGRESSION THE PROPER TOOL OF MODELLING A BEHAVIOUR OF REAL SYSTEMS?

Methodological assumption that multiple linear regression is an adequate tool of modelling the behaviour of real systems is checked. To do this the experiment is organised on the basis of simple “real” system represented as finite discrete automaton. Main result is that in situation of “black box” modelling the approximation of output variables with multiple linear regressions (from several samples and under different conditions) may not fulfil any of criterions of feasible approximation of systems behaviour, also in situations where real relation between input and output variables is strictly linear and only one of variables is omitted.