Chopper Technologies

The model

How we predict when a homeowner will sell

A gradient-boosted classifier that estimates the probability a property sells within twelve months — built to be measured honestly and to explain every score it produces.

0.83

ROC-AUC

Measured on out-of-time validation, not on the training window.

~5M

Training snapshots

Roughly 298,000 properties observed across 17 quarterly cohorts.

19+

Point-in-time features

Equity, tenure, lien position, rate lock, neighborhood turnover, and more.

12 mo

Prediction horizon

Probability the property sells within the next twelve months.

Trained on history, not hindsight

The training set is built from quarterly snapshots of every property in the county. For each snapshot, features are computed looking strictly backward — only transfers, liens, and valuations recorded before that date — while the label looks strictly forward, asking whether the property actually sold in the following year.

That boundary is what makes the accuracy number meaningful. A model that can see the future during training will always look excellent in testing and fail in production.

Only real sales count

Public records are noisy. Properties move between family members, into trusts, and through foreclosure — none of which is a listing an agent could have won.

The model counts only arm's-length market sales, filtering out quit-claim transfers, sheriff's deeds, and nominal transactions. What it predicts is a real transaction, which is the only kind worth calling about.

Every score explains itself

Each prediction returns the five factors that moved it most, expressed as how much each one shifted the probability. An agent opening a contact does not just see 74% — they see that the owner has held the property eight years, that equity crossed 60%, and that homes on the street have started turning over.

A score without reasoning tells you who to call. A score with reasoning tells you how to open the call.

Score tiers

Very Likely
70%+Strong sell signals — high equity, long tenure, distress indicators present.
Likely
50–70%Multiple moderate indicators pointing toward a sale.
Possible
30–50%Some signals present, worth nurturing.
Unlikely
10–30%Few indicators.
Very Unlikely
<10%Stable — recent purchase, low equity, no distress.

Where it runs today

The propensity model is trained and in production on Cuyahoga County, Ohio, where the depth of available deed, mortgage, and assessment history supports the full feature set. We are expanding county by county, and every new market has to clear the same out-of-time validation before a single score is shown to an agent.

We would rather tell you a market is not ready than ship you a number we cannot stand behind.

Ask about your market