Clicks Predict Grades Only Where Courses Are Built for Clicking

Every time a student logs into Moodle, the learning management system quietly records a trail of digital footprints: quiz attempts, assignment submissions, forum posts, file downloads, and the timestamps that bind them together. For more than a decade, universities have hoped that this exhaust of everyday learning could be mined into something precious—an early warning system that flags struggling students before they fail. A new study from researchers at Arab American University in Jenin, Palestine, adds a crucial and sobering twist to that ambition. Published in Cluster Computing, the work by Wala’a Shehada, Huthaifa I. Ashqar, and Ahmed Ewais shows that the predictive power of Moodle engagement is not a universal property of the platform at all. Instead, it depends dramatically on what kind of course a student is taking, and whether the course design pushes students toward the kinds of measurable interactions that algorithms can read.

The research team set out to answer a deceptively simple question: how well does digital engagement actually explain academic performance? To do so, they analyzed three undergraduate courses at their institution with sharply different pedagogical orientations. Research Methods is an assessment-driven, STEM-oriented course in which grades flow from quizzes and structured assignments. Arabic Language is a content-oriented humanities course that relies primarily on lectures and traditional forms of assessment. Palestinian Studies sits somewhere between the two, blending lecture-based content delivery with a moderate assessment component. By comparing these three contexts side by side, the researchers could isolate how course design—not just student behavior—shapes the relationship between clicks and grades.

Methodologically, the study combined two complementary layers of analysis. The first was correlation analysis, examining how strongly different categories of Moodle activity—quiz interactions, assignment submissions, resource views, and temporal patterns of engagement—aligned with final grades. The second was predictive modeling, in which machine learning algorithms were trained on engagement features to forecast student outcomes. This dual approach matters because correlation and prediction are not the same thing. A metric can correlate with grades yet add little predictive value once other features are considered, and a model can achieve high overall accuracy while still failing at the task that matters most: catching the students who are about to fall through the cracks.

The starkest result came from the Research Methods course, where Moodle activity proved to be a remarkably strong predictor of success. Quiz- and assignment-related interactions were highly correlated with final grades, and machine learning models reached predictive accuracy approaching 97 percent. The explanation lies in the architecture of the course itself. Because Research Methods channels its assessment through the platform, the digital trace is not a proxy for learning—it is the learning, or at least the graded expression of it. Every quiz attempt and every submission lands directly in the log data, so the signal the algorithms need is dense, timely, and directly tied to the outcome being predicted. In such environments, learning analytics can function almost like a thermometer placed inside the system it measures.

Yet even in this best-case scenario, the study uncovered a critical weakness: class imbalance. In most real cohorts, failing students are a minority, and machine learning classifiers trained on imbalanced data tend to optimize overall accuracy at the expense of the minority class. The result is a model that is excellent at recognizing students who will pass and comparatively poor at detecting those who will fail—the precise population an early warning system exists to serve. This is not a niche technical quibble. It echoes a growing body of literature on class imbalance handling in educational prediction, and it means that headline accuracy figures can mask a system that quietly misses the very students it was built to protect.

The contrast with the humanities courses was equally instructive. In the Arabic Language course, engagement data showed weak predictive performance, while Palestinian Studies demonstrated only moderate predictive capability. The reasons are structural rather than mysterious. When a course’s decisive assessments happen offline—through in-class examinations, oral performance, or handwritten work—the Moodle log captures only a peripheral slice of student effort. A student might attend every lecture, take meticulous notes, and excel on a final exam without ever generating the volume of platform activity that feeds a predictive model. Conversely, high platform activity in such courses may reflect compliance or navigation rather than mastery. The digital footprint, in other words, is not a neutral measure of engagement; it is a measure of what the course happens to route through the platform.

These findings land with particular force on a debate that learning analytics researchers have been having for years. A widely cited 2016 study by Dragan Gašević and colleagues argued that learning analytics should not promote a one-size-fits-all approach, because instructional conditions shape which engagement indicators matter and how much. The new study provides fresh, department-level evidence for that thesis, showing that the same platform, the same institution, and even the same student population can yield wildly different predictive landscapes depending on disciplinary practice. A dashboard that works brilliantly for an engineering faculty may be nearly blind in a language department—and deploying it uniformly could create a false sense of security precisely where students are least protected.

The practical implications reach into the design of early warning systems and the ethics that surround them. If predictive value is context-dependent, then institutions seeking equity-focused interventions must build models that are sensitive to course design, disciplinary norms, and the structure of assessment, rather than importing generic algorithms across the curriculum. The authors frame their work as a nuanced framework for learning analytics, one that pairs theoretical insight with actionable guidance: audit which activities actually carry grade weight in a given course, match the analytics to that structure, and be explicit about the limits of what log data can reveal. There is also a fairness dimension here. Research on algorithmic bias in education has shown that predictive models can perform unevenly across student subpopulations, and a model whose signal depends on course design may systematically disadvantage students in courses where engagement is legitimately expressed off-platform.

Transparency and reproducibility strengthen the study’s contribution. The dataset and the Python code used for preprocessing, feature extraction, statistical analysis, and predictive modeling are publicly available on Zenodo, allowing other researchers to verify the results and extend the approach to their own institutional contexts. The work also builds on the authors’ earlier research on temporal modeling of LMS logs, and it situates itself within a broad and active field spanning educational data mining, dropout prediction, explainable AI in education, and multimodal data fusion. What distinguishes this contribution is its insistence on context: rather than chasing a single best model, it maps how the explanatory power of engagement data shifts across departmental boundaries.

For university leaders, instructors, and the growing industry of learning analytics vendors, the message is both cautionary and constructive. Digital engagement data is a powerful lens—but only where the curriculum has been designed to pass through it. In assessment-driven courses, Moodle logs can support near-ceiling predictive accuracy and genuinely useful early warnings, provided that class imbalance is handled with care. In lecture-centered humanities courses, the same logs tell a fragmentary story, and institutions that treat them as a universal indicator of achievement risk misreading their students. The future of learning analytics, this study suggests, lies not in bigger models but in smarter alignment: understanding the pedagogy first, then deciding what the data can honestly say about it.

Subject of Research: Predicting student academic performance from Moodle learning management system engagement data across different course contexts

Article Title: Linking moodle engagement to academic performance: context-sensitive insights from departmental and predictive analyses

Article References: Shehada, W., Ashqar, H. I., & Ewais, A. (2026). Linking moodle engagement to academic performance: context-sensitive insights from departmental and predictive analyses. Cluster Computing, 29(12), Article 718. https://doi.org/10.1007/s10586-026-06521-y

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06521-y

Keywords: learning analytics, Moodle, LMS, educational data mining, machine learning, performance prediction, higher education, class imbalance, early warning systems, course design, disciplinary context, predictive modeling