The science
The core unit inside ValueAtlas is not a document or a prompt. It is an observation tied to coordinates and time. Language models are useful as interfaces and reasoners; the durable intelligence comes from structured spatial relationships, temporal sequences, event histories and observed outcomes.
1.4 billion geocoded events: recorded deeds, listings, building permits, transit schedules and ridership, hazard records, employer footprints, retail activity and satellite imagery. Each observation carries coordinates, a timestamp and a source score.
846,000 New York real estate parcels reconciled across 12 record authorities. Matches are time-aware, so a parcel that was split, merged or renumbered keeps a continuous history rather than becoming three unrelated records.
Every property is connected to its schools, transit stops, zoning envelope, comparable sales and hazard exposures. Every edge is validity-dated, so the model can ask what a property was connected to in 2019, not just today.
Developments move through five states — rumored, proposed, approved, built, operational. Markets reprice on announcement and again on delivery, so the engine treats state transitions as first-class price signals.
Repeat-sales indices, hedonic regressions, geographically weighted models, graph neural networks and temporal transformers are combined. Each family is strong in a different regime, and the ensemble weight itself is learned per market.
Every prediction is written to an append-only log with model version, inputs, horizon and confidence band. Calibration is then measured by market, horizon and regime, and published rather than buried.
Accuracy is reported the same way a forecaster would report it: hit rate inside the published band, error magnitude, and whether the direction of travel was right.
See the model applied end to end in the Marina Vallarta case study.