The availability of nutrients exerts major control on ecosystem structure, functioning and responses to global change. Process-based ecosystem models therefore increasingly incorporate nutrient cycles, but forest and other ecosystem modules in integrated assessment models, used to advise policy making based on trade-offs and feedbacks within and among economy, agriculture, forestry etc., nutrient availability is usually poorly accounted for. Here, we explored whether in a statistical random forest model predicting site productivity, replacing soil type by key soil properties (organic layer C:N ratio, upper soil organic carbon concentration (SOC), organic layer pH) would improve predictions across Swedish and European forests. We found substantial variation in the key soil properties and a nutrient availability metric (which à priori integrated the same soil properties), both among and within soil types. Because of the within-soil type variation in nutrient availability, both random forest models using soil properties and these using the nutrient availability metric predicted significantly better forest site productivity than the soil type-fed models across Sweden and Europe. We recommend the inclusion of often available, resource-use related soil properties such as C:N, SOC and pH into random forests that feed into integrated assessment models. Substituting individual soil properties by an à priori defined nutrient availability metric can reduce overfitting in statistical random forest models.