Comparative Analysis of Machine Learning Regression Models for Predicting Mathematics Unified StateExam Scores of Moscow Graduates

In this study, five regression algorithms — Ridge- and Lasso-Regression, Random Forest, Gradient Boosted Decision Trees, and a fully connected neural network — were implemented and compared for predicting 2023/24 Unified State Exam (USE) mathematics (advanced level) scores of Moscow graduates. Models were trained using features that include Basic State Exam (BSE) results across all subjects, achievements in the All-Russian and Moscow school Olympiads, and school grades. The neural network achieved the best performance (MAE = 7.98; MAPE = 12.41% on the validation set; MAE = 8.17; MAPE = 12.63% on the test set). Feature-importance analysis identified the student’s and class’s average BSE math scores, the student’s overall BSE results, and average grades in mathematics subjects as the most influential predictors. These results enable early estimation of a student’s likely USE score and support the development of targeted interventions to enhance final exam performance.