How the Lotvox Turnover Index works

The score sums calibrated points for recorded facts about a parcel and caps at 100. Each point traces to a specific county record, and every surface that shows the score can open that decomposition. Weights are data, versioned per territory, never code; the engine refuses to run a ZIP without a complete, backtested weight set.

It is not a probability and it is not a judgment about an owner. Points are 20 × log₂(lift) per factor, so the total is an indicative multiple of the ZIP's own base sale rate: a score of 20 reads as roughly twice the base rate, 40 as roughly four times. The ladder below is that arithmetic, named. The measured claim is never the band — it is the per-ZIP held-out capture in section 4.

• County and public records only. Assessor rolls, recorder instruments and public-trustee filings. No listing-service data of any kind feeds the score, on any surface.
• Nothing about people. No census attributes, no household characteristics, no purchased "intent" data. Factors describe the property and its recorded instruments — tenure, deed class, entity form, block turnover — never a person's class or state. That is the Fair-Housing posture, by construction.
• Read-only on the public tables. The engine writes only to its own schema; scoring and display only. No outreach is initiated anywhere in it.
• Parcel detail stays behind the territory subscription. Public pages — this one, the accuracy map, the market pages — show coverage and market-level facts only.

Points per factor differ by territory (section 5). Where a county record cannot determine a factor, the factor is absent — rendered "—", never guessed — and a source-map test keeps every factor registered.

Every offered ZIP is backtested point-in-time: parcels are scored as of a past date using only facts knowable then, and the outcome is an arm's-length deed recorded afterwards (quitclaims, death certificates and personal-representative deeds are excluded from the label — a quitclaim is not a listing). The claim graded is top-quintile capture: of the parcels that later sold, what share were already in the top 20% on the scoring date. Random chance is 20%.

A ZIP goes live only when its own held-out capture exceeds 20% on at least 30 held-out sales — the floor below which a capture number is noise. Everything else renders "Score pending" with the reason recorded: 131 ZIPs with too few recorded outcomes yet, 30 that did not beat random, 19 where county records hold no leakage-free history. Never a number worse than a coin flip.

The pilot territory (Broomfield 80020, backtest dated 2026-08-29: features from a January-2025 pre-image snapshot, outcomes Feb 2025 – Jun 2026, 1,743 sellers, base rate 7.15%) inverted the folk heuristic that "long tenure plus high equity equals seller". Measured point-in-time, the sale rate falls with tenure and with accumulated appreciation — the sellers were recent buyers and new construction, not long-tenured high-equity owners:

The two end-points of each ladder are the figures the backtest report quotes; it records the path between them as monotone. A model tuned to the heuristic would have ranked this ZIP upside down.

Scaling to every offered territory (2026-08-30) made the point structural. The same factor name means different things on different county substrates: "no recorded prior deed" is new construction on a snapshot county, rapid churn on Jefferson's 4-deep sale arrays (where it earns ~40 points), and cannot fire at all in counties with 40-year sale chains (0 points, Adams and Weld). Phoenix's dominant signal is purchase recency at twice the pilot's weight; life-event deeds, fixed at 10 in the pilot, measure up to 19–21 points in older Weld and Larimer communities; Douglas clears the gate on block turnover alone. Two Sun City ZIPs failed the gate outright — a retirement-market selling pattern the gate refused to paper over.

So there is no shared classifier. Each ZIP gets a time-split calibration where it has at least 150 calibration outcomes, a county-pooled fit below that (flagged as such), and Maricopa runs leave-one-ZIP-out because its sale dates are month-granular and its affidavit feed lags about two months. Validation is always the ZIP's own held-out outcomes. Weights are then refit on the full window and stored per territory with their honesty label — zip-calibrated, county-pooled or fixed-uncalibrated — and the engine refuses any ZIP without them.

• Modest separation at the top. In the pilot, the first and second quintiles were nearly tied out of sample (4.80% vs 4.84% sale rate); the top-40% vs bottom-60% split is the cleaner line.
• Signal decay. The heaviest calibrated factors burn down as builder close-outs and product waves pass; purchase recency and low appreciation are the durable ones. Weights are recalibrated, not frozen.
• Part of the factor set is not backtestable everywhere. Snapshot counties carry no historical owner or mailing data, so absentee, entity, life-event and trustee factors there ship with fixed, labeled weights. The proof loop exists to replace them with measured numbers.
• One temporal fold per county. One snapshot vintage is one fold, not a walk-forward. Every future pre-image and every night's ledger row extends the evidence.
• Fresh score, stale substrate. Scores recompute nightly, but county rolls advance roughly monthly and owner signals weekly; every run records its feature vintages and the surfaces print them.
• Phoenix lags. Month-granular sale dates and a ~2-month affidavit lag mean Maricopa grades accrue about two months behind Colorado's.

Every night the engine records each territory's top quintile permanently. As county deed feeds advance, a view grades every past run: of the parcels that later recorded an arm's-length deed, the share that were top-quintile that night. The panel is served raw — never smoothed — on the score board and, per ZIP, on the accuracy map.

No post-scoring deed has reached the county rolls yet — the panel says so rather than showing a number it has not earned. The published backtest capture per ZIP stands as the claim until the live grade accrues.

"Scored months before the deed recorded, about one in four of the homes that actually sold in the pilot ZIP were already in the board's top 20% — measured on recorded county deeds, graded nightly in public view."

That is the shape of every claim this score supports: the ZIP's own measured capture, dated, beside the random baseline. Not "predicts 87% of sellers", not a national accuracy figure — the per-ZIP number is the product, and the accuracy map publishes all of them.