Student Success by DesignA practitioner’s guide to theory, data, and AI in student success

5.3 · Technology, Data, and AI Applications

Predictive analytics and early-alert systems

Predictive models estimate which students are likely to stop out, fail a course, or miss a milestone, and route them to outreach.

Short shows the summary, takeaway, and questions.

Georgia State University’s GPS Advising, launched in 2012, monitors more than 800 academic risk factors daily and alerts advisors to milestone-based problems such as registering for the wrong course or a low grade in a prerequisite. The university paired the system with substantial advisor hiring, and has reported higher graduation rates with gaps by race and income largely closed.

How a typical model is built
  1. Historical dataPrior cohorts’ records and outcomes
  2. Outcome & featuresWhat to predict; which variables predict it
  3. Risk scoreA probability for each current student
  4. Threshold & actionWho gets contacted, by whom, how

Design considerations

  • Predict outcomes the institution can influence.
  • Prefer signals (missed milestones, course grades) over fixed traits.
  • Define the intervention before the model; a score with no response is noise.
  • Check accuracy for each student group as well as overall.

Equity and bias risks

  • Including race, income, or ZIP code can encode structural inequity as individual risk.
  • Gándara et al. (2024) found common models more often falsely predicted failure for Black and Hispanic students.
  • Risk labels can steer students away from majors or reduce expectations.
  • Outreach that reads as surveillance can damage belonging.

Theories behind this section

In the Campus Enterprise Model

Where this work sits in campus systems, capabilities, and data.