Distress score formula

The distress score is a 0-100 rollup of six public signals. Each component is capped so that no single source dominates a building's score.

• Complaint velocity: up to 35 points from distinct HPD complaint IDs and allowlisted 311 rows in a 90-day window.
• Violation severity: up to 25 points from open HPD violations, weighted by class (C ≫ B ≫ A).
• Litigation: up to 10 points from active HPD litigation cases.
• Evictions: up to 10 points from residential evictions in the trailing 12 months.
• Speculation: up to 10 points from Speculation Watch flags. Capped at 10 to prevent the small adax-9mitdataset (~800 rows, quarterly cadence) from dominating an otherwise low-distress building's score.
• Rent stabilization weight: up to 10 points reflecting the count of stabilized units at risk.

Score version v1-2026-07-15. Every score row and export CSV carries this version stamp.

What outreach-quiet looks like

A distress score between 1-9 is honest evidence that the building has minimal public signal. This is not an error - it is what a low-signal, healthy-enough building looks like inside the product. When a visitor lands on a building detail with a very low score, the page renders that number and calls it out explicitly rather than hiding the building.

Portfolio linkage & confidence

Buildings are linked into portfolios via Union-Find over an edge graph derived from HPD registration contacts. Two edges exist between buildings if they share either:

• a normalized HPD-registration contact name (owner / head officer / agent), or
• a normalized business address.

Business addresses that appear on more than 50 distinct contact names are excluded before linking. Confidence is min(1.0, edge_density * sqrt(building_count) / sqrt(3))per cluster. Standard confidence ≥ 0.60; low confidence < 0.60. Two-building clusters need corroborating edges to reach standard confidence.

Known false-merge patterns: LLC shells sharing a management-company address; N/A / unknown owner names collapsing unrelated portfolios. Known false-split patterns: same landlord using different registration contact aliases across neighborhoods; portfolios recently changing management.

Supercluster ceiling

Any cluster with more than 1,000 buildings is treated as a supercluster and excluded from ranked outreach queues. The 50-contact high-frequency address exclusion prevents most spurious merges; the ceiling is a hard safety net for the residual over-merges. Supercluster member buildings still appear individually in the intake view with their own distress score.

Priority score & unit exposure

Ranked queues use a priority score, not raw distress. Priority is distress_score × unit_exposure_multiplier × stabilization_multiplier. Unit exposure scales with total residential unit count so that a 220-unit distressed building is prioritized over a 6-unit distressed building. Stabilization multiplier weights buildings whose rent-stabilized unit count is at risk.

311 allowlist

Only the complaint types below are counted from 311 (dataset erm2-nwe9) toward complaint velocity. Version v1-2026-07-15 effective Jul 15, 2026.

Baseline set from REQUIREMENTS.md; expanded with heat/water/pest/safety variants observed in the 2023-present erm2-nwe9 window as agency='HPD' or location_type='RESIDENTIAL BUILDING'. Change-control rule: any addition/removal creates a new version_label (never edits an existing row).

Watchlist comparison

The product compares its top-100 portfolio ranking to the Public Advocate's most recent Watchlist edition. Overlap must be at least 50% for the comparison to pass. Each row where ranks differ is tagged with one of four difference categories:

method_difference

The Watchlist weights only violations per unit per month; this product additionally weights litigation, evictions, and rent-stabilization exposure. Rankings will diverge on portfolios whose distress is disproportionately in one category.

data_recency_difference

The Watchlist edition is annual; our data is refreshed daily to weekly. A distress spike in the trailing months surfaces here first and only appears in the Watchlist at the next annual publication.

universe_difference

The Watchlist restricts to buildings meeting a specific size or violation floor that this product does not apply. Portfolios present here but absent from the Watchlist may consist of buildings below that floor.

linkage_difference

The Watchlist's hand-normalized landlord identity differs from this product's contact-graph portfolio. A single landlord may appear across multiple product portfolios or vice versa.

Overlap threshold >= 50% is honored per REQUIREMENTS.md done-when #5. When overlap falls below threshold, the validation view calls it out honestly and directs the reader to the categorized breakdown.

Score validation

An independent building sample is scored by hand and compared with the product's score. Per-BBL heuristic tier, product score, agreement, and disagreement reasons live in nyc.tenant_outreach_score_validation_sample.

WOW reconciliation

REQUIREMENTS.md requires a Jaccard similarity >=0.70 for at least 8 of 10 fixed BBLs when comparing product portfolio boundaries with the Housing Data Coalition's Who Owns What (WOW) portfolios.

Cluster boundaries can legitimately differ because (1) this product links via Union-Find on HPD registration contact names and business addresses while WOW additionally uses ACRIS buyer/seller deed chains and HPD manager-contact networks (not publicly available as an API); and (2) this product's high-frequency address exclusion (business addresses shared by more than 50 distinct contact names) is an operationalization of WOW's qualitative heuristic, not an exact replica. Mismatches are documented, explainable differences, not errors requiring resolution.

Helpline drop-in fixture

The product accepts a CSV drop-in of aggregated PEU helpline records. The fixture-run aggregate panel on /tenant-outreach/helpline-drop-in shows the latest batch's matched / ambiguous / unmatched / rejected counts and quality status. Thresholds: PASS at >= 85% matched, WARN at >= 75%, FAIL below 75%. No tenant PII enters the product; only aggregate counts and quality metrics are stored.

Prior art & credit

• NYC Housing Data Coalition, Who Owns What (WOW): pioneered contact-graph portfolio linkage from HPD registrations, plus ACRIS deed chains. This product's linkage method builds on and simplifies that approach.
• NYC Public Advocate's Worst Landlords Watchlist: annual violations-per-unit ranking. Used as the primary comparator in the validation view.
• NYC Housing Preservation & Development: owner of the raw HPD Registrations, Registration Contacts, Complaints, Violations, and Litigations feeds.
• NYC Department of Investigation: the Executed Evictions marshal dataset.
• NYC HPD Rent Stabilization roll: the last publicly available stabilized unit counts.