Multi-horizon Student Performance Prediction Using Spatio-temporal Models

Abstract

Predicting academic outcomes early remains a fundamental challenge in learning analytics, particularly when the goal is to identify at-risk students before it is too late to intervene.We introduce a multi-horizon forecasting framework leveraging weekly behavioural and assessment traces from the Open University Learning Analytics Dataset to predict student success at seven temporal checkpoints, from Week 5 to full course duration.A hybrid CNN2D-LSTM architecture is evaluated alongside CNN2D-Only and LSTM-Only variants and four classical baselines, all trained under a strictly time-sliced protocol to prevent temporal leakage, verified through a passive audit and a canary injection test.Results show a consistent improvement with observation length, with all three deep learning variants achieving AUC-ROC above 0.974 at full horizon, averaged across the five ablation seeds and both presentations.Architectural differences between variants are marginal and statistically non-significant, while classical models remain competitive once sufficient data accumulates.Generalizability is confirmed via Leave-One-Module-Out and Leave-One-Presentation-Out cross-validation across unseen cohorts.