Property Priors
Home / Accuracy
The accuracy ledger

We grade our own reads. Here is the ledger.

An estimate never meets an outcome, so an estimate tool can never grade itself. Priors tracks what actually happens: the resale that records, the lease that fills, the calendar that books. That means our reads can be tested against reality, in public, misses included. This page regrades itself every week.

One boundary, stated up front: no number can see inside a specific house. A renovated kitchen, a tired roof, an amenity the neighbors do not have all move the real answer, and they only enter the picture when someone walks the property or tells us what has been done to it. Within that boundary, the grades below are about as close as a number gets.

Panel 1 · rent, graded

The rent read lands within 8.2% on 2,049 leased rentals.

For 2,049 leased Memphis rentals we ran the read blind, then compared it to the rent each one actually leased at. Median miss: 8.2%, with 88% of reads inside 25% of the leased rent.

8.2%
median miss against the rent each listing actually leased at
88%
of reads landed within 25% of the leased rent
31 days
median list-to-lease on the 2,405 rentals tracked to the day they came down

The basis, stated: signed rents are not public record in Tennessee, so each read is graded against the final advertised rent at the moment the listing leased, the closest recorded number to the signature.

Panel 2 · resale, graded

1,358 resale reads, run blind and graded against the recorded price.

For 1,358 finished Memphis flips, we reconstructed the read as it would have printed before each resale, using only records available at the time, with the house's own history excluded. Then we compared the read to the price that recorded.

8.7%
median miss, all 1,358 graded reads
8.0%
median miss on the 1,086 high-confidence reads
84%
of reads landed within 20% of the recorded price
12.5%
what a plain zip-median guess missed by on the same houses

How the grading works: for each realized flip in the 18-month test window, generated Jul 2026, the read is reconstructed as of the resale date using only records available at the time, with the test house excluded from its own evidence, so no read ever sees its own answer. The confidence tier printed on every live read is the same tier graded here: thin pockets miss worse, and the read says so before you rely on it.

Addressthe readrecordedmiss
2634 Sleepy Bend Cv38133 · sold Jul 2026 · 12 comps $225,436the read $235,000recorded 4.1%miss
4835 Dianne Dr38116 · sold Jul 2026 · 12 comps $155,082the read $168,800recorded 8.1%miss
5076 Waters Edge Cv38141 · sold Jul 2026 · 12 comps $277,901the read $290,000recorded 4.2%miss
6727 Bent Tree Ave38134 · sold Jul 2026 · 12 comps $212,513the read $216,000recorded 1.6%miss
486 Brook Ridge Cl38018 · sold Jul 2026 · 11 comps $392,255the read $380,000recorded 3.2%miss
1918 Cherry Rd38117 · sold Jul 2026 · 12 comps $231,963the read $216,500recorded 7.1%miss
3019 Capri Rd38118 · sold Jul 2026 · 12 comps $149,592the read $144,500recorded 3.5%miss
775 E Davant Ave38106 · sold Jul 2026 · 12 comps $127,058the read $139,400recorded 8.9%miss
Recent graded reads, one per zip, high-confidence tier. Addresses are public sale records.

Where we missed

A ledger that only shows hits is an ad. These are recent reads the record disagreed with:

Addressthe readrecordedmiss
2002 Cloverdale Dr38114 · sold Jul 2026 · 12 comps $84,930the read $115,000recorded 26.1%miss
1854 Dearing Rd38117 · sold Jul 2026 · 12 comps $228,229the read $174,000recorded 31.2%miss
Published on purpose. Misses cluster where comps run thin; the confidence tier on every read exists to say so up front.
The loop

This number gets smaller over time, on purpose.

A published grade is only half the discipline. The other half is the loop behind it: every read is graded against what later recorded, and the model only changes when the change makes the grade better. The standing rule is simple. No update ships unless it beats this ledger on the full blind test.

This is the part an estimate tool structurally cannot do: an estimate never finds out it was wrong, so it never gets better on a schedule. Priors does, because it is built on outcomes, and the numbers on this page move as it improves.

Panel 3 · short-term, the method graded

We graded the market-average method against 1,546 real calendars. It misses by 50%.

Every Airbnb projection you have ever seen is a market average: a zip and bedroom-count median dressed up as your number. We tested that method against 1,546 Memphis listings with a full year on their calendar: for each one, the average of its own class, with the listing itself excluded, was asked to guess its year. The typical guess missed by 50%, and landed within 25% only 27% of the time.

50%
median miss of a zip + bedroom average against a real listing's year
27%
of listings landed within 25% of their class average
1,546
full-year Memphis calendars graded in the test

What this means: no average can price a specific house. Two listings on the same block with the same bedrooms can gross far apart on operator quality alone, which is why a market-average projection is a guess wearing a decimal point. The per-address read exists because of exactly this, and grading it against subsequent calendar years is the next panel on this ledger.

Run it yourself

The same graded engine reads any Memphis address.

Run my address, free Go Pro · $29/mo First full read free every month. Pro is unlimited, 7-day trial, cancel in one click.

Every figure on this page is computed from public records and our own tracked observations, regraded on a weekly cycle; this bake: July 18, 2026. Public records contain errors, omissions, and recording lags. Grades describe historical model performance, not a promise about any future read. Nothing here is an appraisal, a broker price opinion, or lending, investment, tax, or legal advice.