The problem with raw test-score rankings
Most public school rankings sort by raw proficiency. That math favors wealthy suburbs over high-poverty schools, regardless of how well either is doing relative to its peers. A school where 70% of students hit proficiency looks great. A school where 40% hit proficiency looks bad. The first might be coasting on a wealthy student body. The second might be one of the most effective schools in its state.
The fix is to control for who the school serves before ranking it. That's what BeatsExpectations does.
How we calculate it
For each US state with sufficient data, we run a per-state regression of average school proficiency on free-and-reduced-lunch share. The regression gives us a predicted proficiency for every school based on its FRL profile alone. The difference between actual and predicted is the school's BeatsExpectations score, in percentage points.
- OUTPERFORMING: top 10% within the state by BeatsExpectations score
- AS EXPECTED: middle 80%
- UNDERPERFORMING: bottom 10% within the state by BeatsExpectations score
Tier thresholds are state-specific because each state uses its own assessment with different rigor and cut-scores. A residual of +12 in California is not the same as a residual of +12 in Massachusetts.
What this rewards
Schools that show up as OUTPERFORMING tend to share characteristics that the research on effective high-poverty schools has identified repeatedly: long principal tenure, structured K-3 literacy instruction, extended instructional time, stable teacher staffing, and a clear curriculum. We track that subset in our separate beating-the-odds analysis.
The table above shows the top 150 outperformers nationally, ranked by raw BeatsExpectations score. The full per-school score is available on every school page on allk12.com that has enough assessment data to compute it.
Limits
The score uses FRL share as the only demographic control. ELL share, special-education share, and the racial composition of the student body also predict outcomes; future versions of the score will incorporate them as state-level reporting catches up.
The per-state regression also assumes that the FRL-to-proficiency relationship is linear within each state. That's a simplification. In states with strongly bimodal FRL distributions (lots of very low and very high FRL schools, few middle), the residuals at the extremes can be noisier than the rest.
BeatsExpectations is a relative measure within a state. A school in the top 10% of one state is not directly comparable to a school in the top 10% of another state, because the underlying assessment is different.
Methodology
Sources: NCES Common Core of Data 2024-25 (FRL share, enrollment, school metadata) and state-native assessments for the most recent available year, typically 2023-24 or 2024-25. Composite proficiency is the average of math and reading/ELA percent-met-or-exceeded across all grades tested at the school. Schools with fewer than four reported assessment rows in 2024 or 2025 are excluded. Virtual schools are excluded because their FRL reporting and assessment participation are not comparable to brick-and-mortar schools. States with fewer than 30 schools meeting the data threshold are skipped.
The regression is ordinary least squares: proficiency = a + b × (FRL share), fit independently for each state. We publish the code that computes the score in the public allk12 repository so anyone can audit or reproduce it.
