Forecasting student academic performance is a basic endeavor in educational data analysis, encouraging early intercession and improved learning strategies. This study analyzes the viability of three machine learning models’ RF, SVM and k-NN Forecasting student performance using time series data. The models were evaluated by residual analysis, feature correlation heatmaps, and error trends. The results illustrate that SVM shown the foremost residual variance, failing to properly capture fundamental patterns. Irregular Timberland and k-NN displayed relatively prevalent execution: however, both models illustrated recurrent error spikes, highlighting difficulties in managing sequential conditions. The leftover relationship heatmap demonstrated analogous error patterns among models, implying a need for advanced feature engineering or ensemble learning strategies.