A skilling platform running a six-month full stack course reported 71 percent monthly active users and treated it as a healthy number. Its actual completion rate was 19 percent. Both figures were correct. The gap between them is where most learning analytics programmes live, and it exists because monthly active is a vanity metric that a student checks once and then abandons.
Dropout is rarely a sudden decision. It is a slow disengagement that the platform can see well before the learner admits it, provided the platform is measuring the right things.
Signals That Lead, Signals That Lag
Completion rate is a lagging metric. By the time it moves, the cohort is gone. The useful work is in finding leading indicators, and the research on large online courses has converged on a fairly consistent set.
- Session gap length. The interval between consecutive logins is the strongest single predictor in most datasets. A learner whose gap crosses roughly twice their established baseline is materially more likely to churn, and the effect is visible within two weeks.
- First-week depth. Learners who complete a meaningful piece of work in the first seven days, not just watch an intro video, complete at multiples of those who do not. The first week carries more predictive weight than any month after it.
- Assessment avoidance. Continuing to consume content while skipping graded items is a reliable distress signal. It usually means the learner has fallen behind and is avoiding the confirmation.
- Video completion ratio, not video starts. Starts measure curiosity. The fraction of a video actually watched, and whether the learner scrubs backwards, measures comprehension.
- Help-seeking that goes unanswered. A forum post or doubt ticket with no response inside 24 hours is one of the sharpest churn correlations in the data, and it is entirely within the platform’s control.
Cohorts Beat Averages
Platform-wide averages hide everything that matters. A 40 percent completion rate can be 65 percent among learners who came through an employer sponsorship and 12 percent among self-paying evening learners, and the intervention for those two groups has nothing in common.
Slice by acquisition channel, by starting skill level from the diagnostic, by device class, and by whether the learner is studying during working hours or after them. In the Indian market the device split alone is worth isolating, because a learner doing a coding course entirely on a phone has a structurally different experience from one on a laptop, and their drop points will cluster around different activities.
Also watch the funnel by content unit rather than by week. If 40 percent of a cohort stops at the same module, the problem is that module, not the cohort. That is a content fix, and it is usually cheap once you can see it.
Intervention Is the Point of the Measurement
Analytics that do not trigger an action are a reporting exercise. The pattern that works is a small number of defined triggers, each with a specific response and a measured effect.
A dormancy trigger at the learner’s baseline gap plus a margin, sending a nudge that references the specific next item rather than a generic reminder. An assessment avoidance trigger that offers a lower stakes practice attempt instead of a warning. A human escalation for high-value learners where an actual call from a mentor is justified by the economics. Each of these should be run as a controlled comparison against a holdout group, because nudge fatigue is real and a badly tuned reminder cadence measurably accelerates the churn it was meant to prevent.
At Invexa, we instrument learning platforms so the engagement data feeds defined interventions with owners and holdouts, since a dashboard that reports dropout after the fact is describing a problem the product could have caught in week two.