We propose a method of data reduction that improves the predictions of correlations obtained from heat exchanger measurements. If we define an ideal heat exchanger on the basis of commonly made assumptions, the two heat transfer correlations corresponding to both sides of the heat transfer surface can be simultaneously determined. A local regression analysis, however, gives a multiplicity of possible correlations corresponding to the given data. The best correlations are obtained from this set by using a global regression procedure. Three methods are evaluated for this purpose: genetic algorithms, simulated annealing and interval analysis. All three perform well, with some differences in accuracy and CPU time. The predictions are further improved by correlating the error that is introduced by the assumptions of the ideal heat exchanger. The heat rate predictions are then improved considerably, giving a good idea of the extent to which these assumptions degrade them.