The signals that arrive before results do

Attendance, submission timing, and system access.

Almost every institution holds data that predicts withdrawal weeks before any result exists. A student who has stopped logging in, who has submitted the last two pieces of work at the final hour rather than early, or whose attendance has become intermittent is telling the institution something, and the data is already collected.

The most reliable single indicator across most settings is a non-submission rather than a poor result. A student who does not submit at all has usually made a decision, consciously or not, and the interval between that and formal withdrawal is where intervention is still cheap.

Using this requires somebody to look at it weekly and act, which is the part that fails. Systems generate reports that go to an inbox and are read when there is time, and there is never time in the weeks when the reports matter most, which are the same weeks everything else is busy.

The arrangement that works allocates the task explicitly to a named person with time protected for it, and treats contacting a flagged student as the job rather than as follow-up. Where it is nobody's specific responsibility it will be done sporadically, which is close to not at all in a system where timing is the whole variable.

It is worth adding that the threshold should be set low enough to generate false positives. A system tuned to flag only students in serious difficulty will flag them too late, and a wasted five-minute conversation with somebody who is fine costs almost nothing compared with the alternative.