Building permits are the earliest concrete signal of development intent. By the time a project appears in market reports, CoStar listings, or news coverage, the permit was filed 6-18 months earlier. But raw permit data is noisy — thousands of filings per month across a metro, ranging from deck replacements to billion-dollar mixed-use towers.
The challenge: how do you separate *systematic development waves* (a neighborhood transforming) from *random maintenance* (scattered re-roofing permits)?
The answer is density-based clustering.
Why DBSCAN Fails (and HDBSCAN Succeeds)
Standard DBSCAN requires a fixed epsilon (neighborhood radius) parameter. In urban geography, this is problematic — a 500m radius that captures a tight development cluster in Manhattan misses a sprawling suburban development pattern in Phoenix. You end up tuning epsilon per metro, per neighborhood, per context.
HDBSCAN (Hierarchical DBSCAN) solves this by working at all scales simultaneously. It extracts clusters by cutting the Minimum Spanning Tree at density-adaptive thresholds, without manual tuning.
What We Feed It
Our permit clustering ingests from the Axiom Locus data platform, which maintains building permits with geocoded locations, zoning change applications, demolition permits, and Certificates of Occupancy. Each record carries a lat/lon, permit type classification, filing date, and estimated project value.
Cross-Domain Integration
This is where Axiom's unified platform architecture pays off. Permit hotspots discovered via HDBSCAN are linked to USPS vacancy data, Census demographics, IRS migration data, BLS employment, FEMA risk scores, EPA environmental justice data, and Transit accessibility.
A hotspot that combines high permit density, population inflow, job growth, and transit access is a fundamentally different investment signal than one driven by single-project speculation.