The Dirty Secret of Property Data

Zillow's Zestimate has a median error rate of approximately 7.9% for off-market homes, according to their own methodology page. In a $400,000 home, that's a $31,600 error band. For distressed properties — the ones investors actually want — the error rate is significantly higher because the model has no training signal for distress.

This isn't a Zillow problem. It's a fundamental property of how real estate data works.

Where the Errors Come From

Real estate data has three structural problems:

1. Temporal staleness. Most property data is updated annually (tax assessments) or at point of sale (transaction records). Between updates, the data decays. In fast-moving markets, a 12-month-old assessment can be 20-30% off market value. The data drifts further for properties undergoing distress — the ones where accuracy matters most. 2. Imputation artifacts. When a data point is missing — say, a property's square footage — vendors don't leave it blank. They impute it. Usually with a neighborhood median. This creates phantom precision: a number that looks exact (1,847 sq ft) but was never measured. Feed that into an AI model and it treats the imputed value as ground truth, propagating the error. 3. Siloed sources. Tax data lives in one database. Mortgage data in another. Insurance claims in a third. Building permits in a fourth. Each source is internally consistent, but there's no cross-validation. A property that sold for $200,000 in county records but has a $350,000 mortgage — that's a contradiction nobody catches because the databases don't talk to each other.

How We Solve It

The Kairos Signal approach is to treat real estate data as a signal processing problem, not a data aggregation problem.

Step one: ingest everything — tax records, deeds, mortgages, liens, permits, insurance claims, utility data, MLS feeds. 19 verticals, each with its own schema and error characteristics.

Step two: don't trust anything individually. Every data point is treated as a noisy observation from one sensor. The "true" value — square footage, assessed value, occupancy status — is inferred by triangulating across sensors. When 5 sensors agree and 1 disagrees, the outlier is flagged with low confidence.

Step three: timestamp everything and measure staleness. Every record carries a last_updated field and a staleness score. Consumers can filter by recency. The system actively flags records approaching their decay threshold for re-ingestion.

Step four: cryptographic hashing. Every enriched record carries a SHA-256 footprint that captures not just the data but the pipeline state that produced it. If a downstream model wants to know whether a specific signal came from a tax record or was imputed from neighborhood data, the footprint tells them.

The Result

Our error rate on distress classification is approximately 0.06 — meaning for every 100 properties we flag as distressed, 94 are genuinely distressed. That's not magic. That's what happens when you treat data quality as a first-class engineering problem instead of an SEO optimization target.

Why This Matters for AI Agents

An AI agent consuming Zillow data to make investment decisions is operating with a ~8% error rate baked into every signal. That error compounds across 100 properties, across 10 metros, across 50 variables. The cumulative error dwarfs any signal the model might extract.

An AI agent consuming Kairos Signal data is operating from verified, cross-referenced, timestamped, footprinted intelligence. The difference in investment outcomes isn't marginal — it's categorical.

---

Kairos Signal: Real estate data that doesn't lie. 500K+ enriched signals, 19-source triangulation, cryptographic verification. Explore data products →