In non-limiting examples of the present disclosure, systems, methods and devices for optimizing machine-learned tree ensemble prediction systems are presented. A plurality of instances may be processed by a tree ensemble. Determinations regarding the expected output values of one or more nodes of the tree ensemble may be made based, at least in part, on the processed instances. Further determinations regarding the node contribution values of one or more nodes of the tree ensemble to downstream nodes may be made based on the node output values. In some examples, feature value ranges may be computed for one or more features of the tree ensemble. One or more tree ensemble optimization operations may be performed based on the information determined from one or more of the above-described operations.