Machine learning models can be used to predict the expected academic performance of students in middle education based on their socioeconomic context. We show how these models can be trained and interpreted to gain insight into the main traits from socioeconomic context that influence academic performance in standardised tests. Our study is presented as a policy brief because the trained models can be used to detect students that outperform their expectations. This allows government agencies responsible for university grant allocation to identify students that should be awarded grants for pursuing tertiary education. We used databases of the Colombian Institute for the Promotion of Tertiary Education (ICFES) as an example of our method.This document has been cited by different academic work in Colombia. Unfortunately, ICFES have removed the public access to the PDF for several of the research results funded by them. We post this PDF in SocArXiv to allow access to our results for future research projects. The document is written in Spanish.