The research evaluates the performance of filter-, wrapper- and embedded-based methods for feature selection adjustments that enhance academic performance predictions from educational data. Through the application of Kalboard-360 (480 students, 16 features) with 649 students in the UCI Student Performance (33 attributes) dataset we tested 10 FS algorithms including Chi-square (Chi2), Correlation-based Feature Selection (CFS) alongside SelectFwe and sequential selection and LASSO and Random Forest importance as embedded methods. The performance evaluation employed RF alongside SVM and DT as classification algorithms. The data showed that wrapper methods specifically SelectFwe (SFWEF) produced the best accuracy rating at 74.3% when used with RF algorithms in agreement with Tariq [1]. The implementation of these techniques proved expensive by requiring 2–3 times longer runtime compared to normal filter method operation. Filter evaluation methods such as GainRatioAttributeEval achieved an optimal trade-off between performance speed-up and predictive accuracy. It improved training efficiency by 15–20% at the same accuracy rate of 76.45% as reported in Enaro [2]. The Chicken Swarm Optimization (CSO)-based FS attained 84.89% accuracy in embedded technique evaluation but faced deployment challenges due to its complexity according to Huynh Cam et al. [3]. This research presents important trade-offs because wrapper methods deliver higher accuracy by using more computational resources but filter approaches provide suitable solutions which work well with real-time requirements. The study strengthens previously established arguments by Sokkhey [4] about the requirement for hybrid FS frameworks in educational data mining (EDM). A thorough evaluation of F1-score and accuracy along with runtime performance enables educational leaders to develop predictive models which align with their specific goals between accuracy and processing speed requirements.