Condominium valuation at county scale

How machine-learning cluster analysis and ensemble modeling produced defensible, market-derived value estimates for thousands of condominiums across Pinellas County, Florida.

The challenge

Condominiums resist conventional mass appraisal. Two units in the same building can differ sharply in value for reasons — floor, view, exposure, association health — that a single county-wide regression averages away. The task was to value thousands of units accurately and, crucially, to explain why each estimate was what it was.

Our approach

We let the market define its own segments before modeling value:

1. Cluster analysis. Unsupervised learning grouped units by the attributes buyers actually price by — waterfront, building era, unit size and amenity profile.
2. Ensemble modeling. Within each cluster, an ensemble of estimators voted on value, steadier and more accurate than any single model.
3. Held-out validation. Every estimate was tested against sales the model never saw.

Findings

Segmenting first roughly halved valuation error versus a single county-wide model, and — because every unit carries its cluster and drivers — each estimate arrives with a rationale that holds up under review.

Figures are illustrative of the study's structure and results. Full methodology and data provenance available on request.