The article deals with the selection of input informative variables for modeling complex, weakly formalized systems. A hybrid method for constructing neuro-fuzzy models with a choice of informative variables is proposed. To select informative variables, a method based on the use of maximum correlation between output and input signals is proposed. Its effectiveness is shown for models that require a lot of computational resources for training. The proposed method is used to find the points of gold occurrences in the Taldan perspective area. It is shown that the choice of informative variables gives an improvement in the accuracy of the model by 10%, as well as a good explanation of which of the geological features make the greatest contribution to improving the accuracy of forecasting gold occurrences.