Next Teardown Check a zip

How the score works

Every single-family house gets a score for how likely it is to be bought and rebuilt. Here is what goes in, what stays out, how it was tested, and where it is weak.

Measured November 17, 2026 on Nashville (Davidson County, TN) with past data, on one time-split test window. Parcel records run through November 3, 2026. It will be re-run on the live pipeline as it ages, and the numbers here will be replaced.

What the score uses

Only facts about the property and its sales: lot size, land value per square foot, house age, house size, land share of total value, the ratio of house value to land value, an estimate of how much more floor area the lot could carry, years since the last sale, and how often that zoning district has seen teardowns in the past. Each house gets short reasons in plain words, such as "built 1951" or "the house is worth half of what the land is".

The model is a regularized logistic regression. That is a plain weighted sum, so every reason can be traced.

What it deliberately does not use

  • Owner names, mailing addresses or phone numbers. None are stored or shown.
  • Anything about the owner's age, or a stand-in for it.
  • Estate, probate or deceased flags, tax relief for the elderly or disabled, family size.
  • Neighborhood demographics.
  • Name-based guesses of any kind.

One yes or no flag, whether the mailing address differs from the house address, was in the model we tested. The earlier audit found it adds little, so we are not carrying it.

How it was tested

We trained on teardowns before a cutoff and tested on teardowns after it, so the score never saw the answers. The headline test window is May 14 to November 13, 2026: 113 houses were torn down among about 57,800 single-family houses. Two earlier six-month windows were also run. A teardown here means a residential demolition permit matched to a new single-family house on the same lot.

Real teardowns inside the top slice of the list, test window May 14 to November 13, 2026, 113 teardowns
MethodTop 1 percentTop 5 percentTop 10 percent
Score (0.878 area under the curve)24 (21.1 times chance)46 (8.1 times)68 (6.0 times)
Simple rule, age and land value (0.823)7 (6.1 times)27 (4.7 times)46 (4.0 times)

One time-split window. The score's top 1 percent lift was 15.8 and 16.1 times in the two earlier six-month windows. Chance would put about 1.1, 5.7 and 11.3 teardowns in the three slices.

Precision, in plain terms

24 of about 578 houses in the top 1 percent were torn down in the window tested. That is a little over 4 percent, so about 96 in 100 houses near the top of the list were still standing. The list is ranked by likelihood. It is not a prediction that any house will be torn down, and it does not say which one will go.

The top 1 percent figure rests on only 24 hits. Its range, from a bootstrap of the test set, is 14.0 to 29.0 times chance, so read 21.1 as a rough size, not a promise. The top 5 and 10 percent figures rest on more hits and are steadier.

Against the rule you already use

A simple rule built from age and land value is the fallback score. It caught 46 of the 113 in its top 10 percent, against 68 for the model. The model clearly wins only at the top 1 percent (24 against 7), and a builder who knows an area already has most of that rule in their head.

What the checks found

  • Zoning carries much of the top 1 percent. Without the zoning feature its lift is 7.0 times chance, not 21.1.
  • 736 of the 1154 houses in today's top 1 percent (64 percent) sit in a one-acre zone that the default filter hides. A fairness check on neighborhood demographics was run, and its method is still open with a lawyer.
  • The score was trained on 2024 property records and applied to today's, so house ages have drifted by about two and a half years.
  • The structure fields come from the county's current building layer, which only holds houses that still stand.

Negative controls

With the labels shuffled, the score should fall to about chance. Shuffling the training labels gave 0.18 times chance at the top 1 percent, and shuffling the test labels gave 0.53 times (chance is 1.0). Those runs are noisy because the top 1 percent holds few teardowns. The test can fail, and it does when it should.

Limits

  • Sales reach the county file 2 to 4 weeks after closing. Parcels and sales run through November 3, 2026, so a house that sold, was rebuilt or was split in the last month can still show. We hide any with a later sale, demolition or new-build permit we can see.
  • Teardowns on parcels with no usable building record are left out of scoring, so recall over all teardowns is lower than the table suggests.
  • The top of the list leans toward large-lot estate zoning and a few zips. Many top houses were on multi-acre lots, which is why you set your own filter.
  • Matching a demolition to a new house is imperfect. 63 percent of 1,731 residential demolition permits were matched. A match by map location can pair a neighbor's new house, and these matches have not been checked by hand.
  • Nothing checks that a house is livable or can legally be demolished or rebuilt. Historic overlays, flood areas and easements are not scored.
  • Tested on Nashville only. The result may not hold in another county.

Not yet done

The test has not been repeated on dated snapshots of the assessor's data, which is what a clean re-test needs. Only three time windows were run. Reasons text has not been reviewed by a second model or a builder. We will re-run all of this as the live data ages and replace the numbers here.

Questions on any of this: ask us or read the zip pages for the permit counts behind the maps.