How the score is built and validated.

Direct from the county.

Every night the pipeline pulls from the source-of-truth system for each of Indiana's 92 counties. For most counties that's the assessor's GIS portal (parcel polygons, owner of record, mailing address, assessed values) plus the recorder's sales-disclosure database (deeds, transfer dates, sale prices). Nothing is scraped. Nothing is licensed from a national aggregator whose data goes stale by the quarter.

Files arrive in a mix of formats — ArcGIS Feature Services, raw CSV exports, state Sales Disclosure Forms. The first stage normalizes them to a canonical schema keyed by a single parcel ID per county.

Assessed values, deeds, and geometry on one record.

The normalized assessor rows join to deed history by parcel ID, to zoning polygons by spatial containment, to FEMA flood overlays by the same, and to owner-portfolio rollups by an exact match on the owner string.

Owner strings get classified into six entity types — Individual, LLC, Corporation, Trust, Partnership, Government — using a curated pattern matcher. The classifier accepts gray areas (e.g. "Smith Family Trust" vs "Smith Family LLC") explicitly rather than guessing.

Each signal is a deterministic rule.

Signals are not machine-learned in a way that would make them inscrutable. Every signal is a small deterministic rule — an absentee owner is one whose mailing address differs from the property address after normalization; a long-hold parcel is one where the most recent recorded transfer is more than ten years ago; a tired-landlord flag fires when the same owner appears on three or more absentee parcels.

The predictive power comes from the signal catalog and the weights — not from a black box. A reader with the rules can reproduce any score.

Calibrated against actual closings.

The weights on each signal came from regressing the binary outcome "did this parcel have an arm's-length sale in the trailing twelve months?" against the signal vector across a corpus of ~48,000 Indiana parcels. Signals that showed high lift kept their high weights; signals that showed no lift were cut from the catalog entirely.

• Estate / heirs: closing rate 6.4× base rate — weight 30
• Sheriff's sale in history: 4.8× — weight 20
• Absentee + long-hold compound: 3.1× — weights stack
• Recent quitclaim alone: 1.8× — weight 8

A simple sum, capped at 100.

The raw score for a parcel is the sum of its active signal weights. A parcel can "top out" — the sum is capped at 100 — once it accumulates enough signals. We don't re-weight or normalize at the tail; the score remains an interpretable sum.

Parcels are binned into three tiers at display time: cold (0–34), warm (35–69), and hot (70+). Tiers are display-only — you can filter, sort, and alert on the raw numeric score.

Measured every quarter.

Each quarter we re-run the calibration on the trailing four quarters of MLS closings and compare the predicted score distribution to the observed closings. The weights are adjusted when and only when the regression shifts by more than a standard error on any individual signal.

When we bump a weight, we publish the change. The accuracy page surfaces a running record of score-to-closing agreement, county by county.

Every score is explainable.

Click any parcel, open the property report. The lead-score card lists every signal that fired on this parcel with its weight. If a parcel scores 78, you can see exactly why — long-hold (25), absentee (20), trust-owned (18), recent quitclaim (8), active sale history (5), land-heavy (2). The math is the math.

That transparency matters because a real estate professional will need to defend the lead to a client. You don't want to be explaining a number you can't verify.