Predicting Students’ State Examination Results based on Previous Grades and Demographics

Abstract

All Estonia’s upper-secondary school students have three mandatory final state examinations – in mathematics, Estonian language and foreign language. We used data obtained from one school in order to predict student’s final examination scores based on previous grades and demographic data. Our aim was to find most important factors that contribute positively to final results and vice versa. Machine learning package Weka was used to build predictive models and with all tested attribute sets we got accuracy over 80% on classification. On continuous models we got mean absolute error close or less to 10, when range of test result was 0-100 points. Most of our attributes were subject grades on scale 1-5, and therefore, as we limited our testing sample with one school only, it gave interesting insights into how some subjects and teachers do contribute to final test results. And, surprisingly, it turned out, that for some test result, most significant predictor was negatively correlated to the result. It means that having bad grades in some subjects was a good predictor of success in final test. Besides being a useful tool for students who want to estimate possible test results beforehand, it is also a useful tool for measuring teachers’ contributions and relationships between different subjects.