The proof
Why trust these numbers?
You don’t have to take our word for it. This page shows the test we’re graded on, how we made sure the test was fair, and where our model still falls short.
We use the official test. Assessors nationwide are graded with the same ratio-study standards we apply here. We did not invent this test.
We test the hard way. Our model is graded only on sales it never saw. It is like grading a student on questions they never studied.
We publish our misses. Where the model falls short, this page says so.
The official test, head to head
The IAAO ratio study compares values to what homes actually sold for, using mortgage-financed sales, which is the standard's own market-value definition. Same homes, same sales, official scoring.
Is the overall level right?
Median ratio · acceptable: 0.90 to 1.10
Do similar homes get similar values?
COD (uniformity) · acceptable: ≤ 15
Are cheap homes over-valued vs expensive ones?
PRD (vertical equity) · acceptable: 0.98 – 1.03
Same question, the preferred test
PRB (vertical equity) · acceptable: within ±0.05
The bottom line: our model passes the level test and both fairness-across-price tests; the city’s roll fails every one. Both were scored on the same homes and the same sales.
What these tests mean, in plain words
“Median ratio” asks whether values are centered on real prices. “COD” asks whether similar homes get similar treatment. “PRD” and “PRB” ask the fairness question that matters most: are cheaper homes valued too high relative to expensive ones? That pattern, cheap homes over-assessed, means lower-income owners quietly pay more than their share of tax.
How we made sure the test was fair
- The model never sees the answer key. We grade it only on sales that happened after everything it learned from. That is the hardest version of the test, and the one that matches how assessments really work.
- The city’s number is never an input. Our model can’t copy OPA’s value, because it never sees it. Otherwise the comparison would be circular.
- Two methods must agree before we flag your home. A property is only marked “may be too high” when two independent statistical methods both put the city’s value outside the likely range.
- No people-data, ever. Race, income, and anything about who lives in a home are never inputs. We use them only afterward, to check the model for neighborhood bias.
What we found in the city’s favor
Two findings came out on the city’s side.
- OPA is not gaming its numbers. A known trick called “sales chasing” means quietly matching assessments to recent sales so the official study looks good. We found no evidence of it in Philadelphia. We ran the standard detection test across two tax years, and it came back clean each time.
- This is a national problem, not a Philadelphia scandal. Research covering roughly 26 million U.S. sales finds the same pattern almost everywhere: the cheapest homes are assessed at roughly twice the rate of the most expensive, relative to what they sell for. Philadelphia sits inside a structural, nationwide failure. A better model moves it, but nothing fully fixes it yet.
What we can NOT claim
- Nobody passes the unfiltered test, including us. The official standard excludes foreclosures, cash-market distress sales, and extreme cases. Scored on every sale with no exclusions, the city’s roll fails badly and our model fails too, by a much smaller margin (see the full table below). “Passes the official test” and “fair to every neighborhood” are different claims.
- Our model is still imperfect on cheap homes. Homes that sell for cash in disinvested neighborhoods are the hardest to value, for the city and for us. Our estimates there carry wider ranges, and we show those ranges instead of hiding them.
- One home can still be wrong. These are statistics. That’s why your report shows the facts to check, not just a verdict.
For the technical reader
The full numbers, definitions, and how to reproduce every figure on this page.
Full table: official IAAO basis (time-adjusted, 3×IQR-trimmed)
| Statistic | Acceptable | City (OPA) | Our model |
|---|---|---|---|
| Median ratio | 0.90 to 1.10 | 0.92 | 1.00 |
| COD (uniformity) | ≤ 15 | 23.0 | 18.6 |
| PRD (vertical equity) | 0.98 – 1.03 | 1.065 | 1.021 |
| PRB (vertical equity) | within ±0.05 | -0.056 | 0.007 |
Full table: every arms-length sale, no exclusions
| Statistic | Acceptable | City (OPA) | Our model |
|---|---|---|---|
| Median ratio | 0.90 to 1.10 | 0.983 | 1.037 |
| COD | ≤ 15 | 34.5 | 24.9 |
| PRD | 0.98 – 1.03 | 1.190 | 1.086 |
| PRB | ±0.05 | -0.234 | -0.084 |
| Typical error (MAPE) | no target | 34% | 26% |
Neither passes; ours fails by a fraction of the margin (PRB −0.06 vs −0.23, COD 25.9 vs 34.5). The gap between the two tables is the 3×IQR trim removing the cash/distressed tail. That exclusion is built into the official standard itself.
Definitions
- Sales ratio
- Assessed value ÷ sale price. 1.00 means the assessment matched the market.
- COD: coefficient of dispersion
- Average % spread of ratios around the median. Lower = similar homes treated alike.
- PRD: price-related differential
- Above 1.03 means cheaper homes carry relatively higher assessments (regressive).
- PRB: price-related bias
- The preferred regressivity test: % change in ratio as value doubles. Negative = regressive; −0.234 means ratios fall ~23% per doubling of value.
Source: out-of-time test slice of model run 20260707T030251Z-baseline, n = 19,484 arms-length Philadelphia sales (regenerated 2026-07-07 via fair-measure export-web-stats); the sale-chasing check uses the assesspy implementation across TY2025–2026. Every figure is reproducible from the open-source pipeline (fair-measure train-baseline, fair-measure ratio-study) and documented in the technical model documentation, including the measurements that did NOT work.
Next: how the model works · or check your own home.