The last decade has witnessed of increasing the application of random forest (RF) models that are known as an exhibit good practical performance, especially in high-dimensional settings. However, on the theoretical side, their predictive ability markedly remains unexplained, especially in coal preparation. RF as a predictive model can tend to work well with large dimensional databases and rank predictors through its inbuilt variable importance measures. In this study, relationships among ultimate and proximate analyses of 6339 US coal samples from 26 states with gross calorific value (GCV) have been investigated by multivariable regression (MVR) and random forest (RF) models. RF method has been used for the variable importance. Models have shown that the ultimate analysis parameters are the most suitable estimators for GCV and that RF can predict GCV quite satisfactory. Running of the best arranged RF structures for the input sets and assessment of errors have suggested that RF models are suitable for complicated relationships.