The random forest (RF) algorithm was used to develop two models for predicting the first-year corrosion losses (C 1 ) of carbon steel in open air in various regions of the world. The first RF model built using combined databases of international programmes ISO CORRAG, MICAT and ECE/UN and tests conducted in Russia is intended for estimation of C 1 in various types of atmospheres in various regions of the world. The second RF model enables the prediction of C 1 in continental areas of the world. The accuracy of C 1 predictions by the two RF and two dose-response functions, i.e. the function presented in ISO 9223 standard and the new version for a non-marine atmosphere, was compared. The reliability of the two RF models was shown to be significantly higher than that of the dose-response functions with exception of the predictions for corrosion losses of carbon steel in regions of Russia with a cold climate.