In short
- Risk is calculated from logins, submissions and grades in the LMS, compared with the student's group.
- It does not use socioeconomic data or how much the student uses the Pharis tutor.
- It is a logistic regression: each signal has a fixed weight and every alert states its reasons in words.
- It is trained on public data from other universities, so it ranks students by signs of disengagement; it is not a calibrated probability for your institution.
- The alert goes to a person, who decides whether and how to act. No decision about a student is automatic.
What data it uses
Pharis reads the LMS every day in read-only mode. For each student, in each course, it measures:
- Activity compared with their group: how many days they logged in to the course in the last 7 and 28 days, compared with their group's median, and how many days since their last login.
- Submissions: what share of assignments already due was not submitted, over the last 4 weeks and since the course started, and how many were submitted late.
- Grades: the average of what has been graded so far, or whether there are no grades yet.
Comparing against the group matters: logging in twice a week can be normal in one course and a warning sign in another.
What it leaves out, on purpose
- Tutor usage. A student who asks few questions may understand everything, and one who asks many may be very engaged. Questions go to the professor's map of doubts, not to the risk signal.
- Socioeconomic and demographic data. The model does not use age, gender, scholarships or income. The only possible financial signal is optional and kept separate (see below).
What it was trained on
There are two models, and the system picks one per student:
- Weekly model. Used once there are at least 14 days of login history. It was trained on the Open University Learning Analytics Dataset (Kuzilek, Hlosta and Zdrahal, 2017): 32,593 students across 22 courses at a UK distance-learning university, observed week by week. It was validated by leaving out one whole course at a time and measuring on that unseen course.
- Grades model. Used while there is not enough login history. It looks at courses enrolled, passed and assessed in recent terms, and it was trained on the public Predict Students' Dropout and Academic Success dataset (UCI), with 4,424 higher-education students in Portugal.
Both are logistic regressions: classic statistical models in which each signal adds or subtracts with a known weight. We chose this type of model because it lets us explain every alert.
From score to alert
The result is a score from 0 to 100. In the weekly model, the level is high from 70 and medium from 55; in the grades model, medium starts at 40. Every alert comes with:
- the reasons in words, for example "has not logged in to the course in 9 days";
- a suggested action, for example "schedule 10 minutes with this student this week";
- how much evidence is behind it, in weeks of data.
If there is not enough data, the student is shown as "no data". Pharis never makes up a number to fill the gap.
The optional financial signal
If the institution shares outstanding balances or late payments, a signal from the last 30 days adds a fixed adjustment to the score: 15 points by default, configurable per institution between 0 and 50, capped at 100. It does not change the model and is not used to train it. Scholarships are recorded but do not add points. Without academic history, a financial signal alone does not produce a score.
Why we do not publish an accuracy figure
Both models were trained on students from other institutions, other countries and other calendars. An accuracy figure measured on that data would say nothing about your institution's students, so we do not present one, either to professors or in the panel.
What adjusts the model to an institution is its own outcomes: who stayed and who left. With at least one full term closed, and around a hundred students with a known outcome, the model and its cutoffs can be recalibrated to that institution.
Who decides
Risk is a signal for a person. The professor, coordinator or relevant area decides whether to contact the student, how and when. Every contact is logged, and at the next measurement Pharis compares the risk before and after to learn which actions work.
Want to see it with your institution's data? Book a demo.