The Additive-Logistic Admissions Model

The model in one sentence

AdmitGPT converts your academic metrics into z-scores against each school's own admitted-student distribution, adds an extracurricular spike score and contextual modifiers, then passes the sum through a logistic function to produce a probability between 0% and 100%.

Step 1: Academic z-scores

Your SAT and GPA are converted to z-scores — standard deviations from the mean — using each college's own admitted-student distribution. This means a 1500 SAT at a school where the average is 1300 is weighted differently than a 1500 at a school where the average is 1480. The combined academic z-score (capped at [-4, 4]) is multiplied by 1.5 in the logit, reflecting that academic strength is the single strongest predictor of admission.

Step 2: Extracurricular spike score

The extracurricular spike is scored across six dimensions:

1. Tier — from local (tier 5) to international (tier 1)
2. Level — from participation to leadership
3. Rarity — how common the activity is among applicants
4. Institutional strength — how well-regarded the activity is by admissions offices
5. Cognitive load — intellectual depth required
6. Validation — external recognition (awards, publications, etc.)

The spike contribution is capped at ±2.0 in the logit so no single activity can overpower weak academics. This cap reflects the reality that even the most impressive extracurricular cannot compensate for a GPA that is significantly below a school's median.

Step 3: Contextual modifiers

Three additional modifiers adjust the logit:

• Intended major fit — competitive majors (CS, engineering, pre-med) receive a negative adjustment; less competitive majors receive a positive adjustment
• International status — international applicants at need-aware schools receive a negative adjustment reflecting the additional competition for limited international spots
• Early Decision timing — ED applicants receive a positive adjustment reflecting the historically higher ED acceptance rates (2–4x Regular Decision)

Step 4: Logistic conversion

The final logit value is converted to a probability using the logistic function:

The result is a probability between 0% and 100% for each school. The model does not use machine learning or neural networks — it is a transparent, interpretable statistical model with published coefficients.

Calibration and accuracy

The model is calibrated on 1,122 self-reported applicant profiles (2020–2023) and validated against the IPEDS college database. Key metrics:

• Ordinal AUC: ~0.74 — the model correctly ranks similar applicants in the right order about three-quarters of the time
• Well-calibrated for 25–75% admit rates — predictions closely match observed outcomes in the middle range
• Systematically under-predicts at sub-10% schools — predicted ~1.3% vs observed ~24.6% in the hardest decile

The under-prediction at highly selective schools is a known limitation of additive logistic models. We report this openly rather than hiding it. For detailed calibration analysis, see the Transparency page.

What the model does NOT capture

No model can capture everything. AdmitGPT does not model:

• Essay quality and personal statements
• Letters of recommendation
• Interview performance
• Demonstrated interest
• Holistic reader judgment
• Institutional priorities (development cases, athlete recruiting)

These factors are significant, especially at the most selective schools. That is why the model under-predicts at sub-10% schools — the published admission rate blends all of these factors, while the model only captures the measurable ones.

Academic basis

The additive-logistic approach is grounded in published admissions research:

• Giani & Walling (2020) — "Access and Outcomes of US Transfer Students" — used logistic regression to model admissions outcomes
• Lee, Kizilcec & Joachims (2023) — "Modeling College Admissions" — applied additive logistic models to admissions data

AdmitGPT adapts this approach for a web-based calculator, adding extracurricular scoring and contextual modifiers that go beyond what academic papers typically model.

Open source

Every formula, weight, and coefficient is published in the open-source repository. You can verify the math yourself, fork the project, or contribute improvements.