Grade progression
Each cohort is rolled forward grade by grade using progression ratios measured from the district's own recent transitions — capturing retention, migration, and public/private shifts as they actually occur locally.
The headline result
Measured against the two easy answers.
A forecast is only worth paying for if it beats what a spreadsheet gives you for free. We hold ours against two free baselines: same enrollment (next decade looks like today) and trend extrapolation (draw a line through the last five years). At every horizon from one to ten years, the District Foresight model is more accurate than both — and the advantage grows with the horizon, exactly where planning decisions are hardest.
District Foresight
Trend extrapolation
Same enrollment
View data table
| Horizon (years) | District Foresight MAPE | Trend MAPE | Same-enrollment MAPE | Projections scored |
|---|---|---|---|---|
| 1 | 1.84% | 2.16% | 2.10% | 16353 |
| 2 | 2.81% | 3.24% | 3.39% | 15855 |
| 3 | 3.75% | 4.34% | 4.65% | 15358 |
| 4 | 4.68% | 5.48% | 5.92% | 14860 |
| 5 | 5.56% | 6.69% | 7.18% | 14362 |
| 6 | 6.49% | 7.96% | 8.39% | 13864 |
| 7 | 7.45% | 9.30% | 9.63% | 13366 |
| 8 | 8.47% | 10.78% | 10.91% | 12869 |
| 9 | 9.40% | 12.29% | 12.16% | 12372 |
| 10 | 10.30% | 13.84% | 13.46% | 11875 |
Ten years out, the model's mean error is 10.3% against 13.8% for trend extrapolation and 13.5% for assuming enrollment stays flat — a 25.6% error reduction versus the trend line. In head-to-head comparisons on identical district-year pairs, the model beats the same-enrollment baseline in 61.6% of 10-year projections. Notably, the trend line itself barely improves on doing nothing — and is worse than doing nothing at long horizons. Extrapolating recent direction is not a forecast.
Calibration, not just accuracy
Know how much to trust a number before you plan on it.
The same backtest that scores the model also calibrates it. The band below is the distribution of errors a district should actually expect at each horizon — the honest planning range around any point forecast we publish.
View data table
| Horizon (years) | Median error | 25th percentile | 75th percentile |
|---|---|---|---|
| 1 | 1.29% | 0.61% | 2.37% |
| 2 | 2.08% | 0.96% | 3.70% |
| 3 | 2.84% | 1.30% | 4.96% |
| 4 | 3.56% | 1.65% | 6.23% |
| 5 | 4.22% | 1.96% | 7.40% |
| 6 | 4.98% | 2.29% | 8.63% |
| 7 | 5.76% | 2.66% | 9.87% |
| 8 | 6.52% | 3.03% | 11.26% |
| 9 | 7.15% | 3.37% | 12.63% |
| 10 | 7.89% | 3.71% | 13.83% |
Half of all 10-year projections land within 7.9% of the eventual enrollment. A district planning a boundary change or a capital program can carry that range directly into its scenarios — and because the range is measured rather than asserted, it is defensible in a board meeting.
How the model works
Cohort survival first, housing signals on top.
The engine is the same family of model state education agencies trust — grade progression (cohort survival) — wrapped in a discipline most forecasts skip: every component is fit and re-fit strictly walk-forward, so no projection ever uses information that was not available on the day it would have been made.
Entry grades
Kindergarten — and any grade that enters the district from outside, like grade 7 in a regional system — is projected from its own recent cohorts; birth and demographic data extend this in engagements.
Housing signal
Building permits for the district's own municipalities enter as deviations from the district's norm — unusual construction booms add students on a measured, multi-year schedule.
Walk-forward fitting
Every regression is re-estimated using only outcomes already observed at the projection date. The backtest is a genuine simulation of forecasting in real time, not a curve fit to history.
The housing signal, measured
New homes add students on a schedule you can plan around.
Fitting the model across 499 districts and three decades yields a measured answer to a question planners usually handle with rules of thumb: how many public-school students does unusual new construction add, and when?
Single-family units
Multi-family units
View data table
| Years after permits | Students per excess single-family unit | Students per excess multi-family unit | Observations |
|---|---|---|---|
| 1 | 0.055 | 0.045 | 16353 |
| 2 | 0.101 | 0.089 | 15855 |
| 3 | 0.151 | 0.130 | 15358 |
| 4 | 0.208 | 0.161 | 14860 |
| 5 | 0.273 | 0.206 | 14362 |
| 6 | 0.320 | 0.244 | 13864 |
| 7 | 0.384 | 0.321 | 13366 |
| 8 | 0.470 | 0.387 | 12869 |
| 9 | 0.569 | 0.465 | 12372 |
| 10 | 0.684 | 0.527 | 11875 |
The yield curve rises steadily — homes take years to fill with school-age children — and single-family units consistently out-yield apartments, in line with the residential demographic multiplier literature. Two honest findings from the same analysis: a district's own grade-progression ratios absorb most of a housing boom within a few years of its start, so the explicit permits term matters most early in a boom; and most of the model's long-horizon edge over plain cohort survival comes from continuous recalibration against observed outcomes. We publish that decomposition rather than let the housing story take credit it did not earn — the full detail is in the PDF report.
Robustness: the model's advantage holds in every era of the backtest — through the 1990s expansion, the mid-2000s housing boom, the 2008 crash, and the post-2010 enrollment decline. Era-by-era tables are in the full report.
Data & reproducibility
Public data, auditable pipeline.
Every series in the backtest is public and citable — a district can audit any number we publish back to a federal file.
- District enrollment by grade, 1986–2024: NCES Common Core of Data (via the Urban Institute's harmonized files).
- Residential building permits by municipality, 1980–present: U.S. Census Bureau Building Permits Survey.
- District–municipality geography: NCES EDGE school district geographic relationship files.
- Pennsylvania October-1 enrollment and official district projections: Pennsylvania Department of Education.