Intelligence tests are used in various scenarios in order to assess individuals' cognitive abilities. As these tests are typically resource-intensive and quite lengthy, we propose a predictive analysis paradigm with the aim of most effectively predicting IQ scores and thus shortening and optimising tests by identifying the most predictive test components. Using the Berlin Intelligence Structure Test for Adolescents (BIS-HB) as an example, we apply machine learning models and successfully predict IQ scores at the individual level. In addition, we identify non-significant and potentially redundant tasks and items and exclude them from the analyses, while maintaining the same satisfactory predictive results. A new direction of research in this area will allow not only the inductive optimisation of intelligence tests, but also the improvement of knowledge and understanding of intelligence in general.