How the PropertyTensor AVM works
Our Automated Valuation Model (AVM) returns an estimated market value for a property together with a calibrated prediction interval and a confidence score. This page explains the modeling approach, how the value is produced, and how we quantify uncertainty with a Forecast Standard Deviation (FSD). It is a methodology overview, not an implementation guide.
Ensemble in log-price space
A blend of gradient-boosted tree models trained on historical arm's-length sales, targeting the logarithm of sale price so errors stay proportional across price levels.
Honest, calibrated uncertainty
Every estimate ships with a prediction interval fit as a heavy-tailed distribution and calibrated so that the stated coverage matches realized outcomes.
Leak-free evaluation
Accuracy is measured with walk-forward validation on future months only, so reported numbers reflect real deployment rather than optimistic in-sample fits.
How we value a property
The point estimate is produced by an ensemble of gradient-boosted decision-tree models. Each model independently learns how a property's attributes and local market map to price; their predictions are then combined into a single blended estimate. Using several complementary learners rather than one is deliberate — each captures different structure in the data, and the blend is consistently more accurate and more stable than any single model.
Models are trained against the logarithm of sale price rather than raw dollars. This keeps error proportional: a fixed percentage miss on a $300K home and a $1.5M home is treated as equally important, which prevents a handful of expensive properties from dominating the fit and makes the error distribution far more symmetric. The final dollar estimate is the median of the fitted distribution, transformed back out of log space.
The returned value is this median, and the accompanying value_low / value_high band and confidence_score come from the distributional layer described below.
What the model sees
The model draws on several families of signal for each property. It does not rely on any single input; the strongest predictors are consistently the location and neighborhood signals rather than the raw physical attributes.
- Physical attributes. Beds, baths, living area, lot size, year built, stories, and structural / amenity characteristics.
- Location & neighborhood.Subdivision and neighborhood price levels, assigned schools, and geographic position — typically the most influential signals in the model.
- Comparable sales. Statistics summarizing recent nearby sales of similar homes drawn from public transaction records, computed only from sales that closed before the valuation date.
- Market & timing context.Local supply-and-demand conditions and broader macroeconomic indicators, each aligned to the property's valuation month.
- Property imagery. Visual signals distilled from non-MLS property and street-level imagery, used to capture quality and condition that structured fields miss.
Market and timing signals are always attached to a property using only information that was available at its valuation date, so the model never sees the future when scoring the present.
Training & validation
Accuracy is measured with walk-forward temporal validation: the model is trained only on sales that closed before a given month and then scored on that month's sales. This mirrors production, where the model must value today's homes using only the past. It is deliberately stricter than a random train/test split, which would leak future price levels into training and report an optimistic number.
We report accuracy in this leak-free regime, and monitor it across many consecutive months to confirm the model is stable rather than tuned to a single lucky period.
Accuracy
Representative off-market performance (no current asking price used), measured with walk-forward validation. Median error is the median absolute percentage difference between the estimate and the eventual sale price.
| Price tier | Median error | Within 10% | Within 20% |
|---|---|---|---|
| Overall | ~6.5% | 70% | 89% |
| $200K – $400K | 4.7% | 76% | 93% |
| $400K – $600K | 5.1% | 76% | 93% |
| $600K – $800K | 5.8% | 71% | 90% |
| $800K – $1M | 7.6% | 62% | 87% |
| $1M – $2M | ~9% | 53% | 81% |
| $2M+ | ~16% | 33% | 61% |
Accuracy is strongest in the high-volume mid-market and degrades at both extremes: luxury homes carry unobservable quality (finishes, views, prestige) and the very bottom of the market includes distressed and non-arm's-length transactions. Measured in the same leak-free, off-market regime, this median error is comparable to — and in our coverage areas modestly better than — published off-market benchmarks from major consumer AVMs.
FSD & the confidence score
A point estimate alone is not enough — a $500K estimate that could plausibly be anywhere from $450K to $550K is very different from one pinned to $490K–$510K. Like every major AVM, we express this with a Forecast Standard Deviation (FSD): a per-property, proportional measure of how far the eventual sale price is likely to fall from the estimate. The confidence_score we return is derived from it, and the confidence label maps to standard industry bands:
| Confidence | FSD range | Interpretation |
|---|---|---|
| High | ≤ 13% | Dense comparable support; anchor tightly. |
| Medium | 13% – 20% | Reasonable support; verify with specific comps. |
| Low | > 20% | Thin or unique market; treat as a starting point. |
Rather than emit a single point and bolt on a separate error guess, our model predicts a full probability distribution over the sale price. We model it as a heavy-tailed (Student-t) distribution in log-price space. This choice is grounded in the data: real valuation errors are close to symmetric but have fatter tailsthan a bell curve — sale prices land far from the estimate more often than a normal distribution would predict. A heavy-tailed model captures those tails honestly instead of understating them.
Crucially, the width of each property's distribution is nota fixed percentage. It reflects how much genuine evidence exists for that specific home — driven chiefly by the availability of high-quality comparable sales and by how much the underlying models agree with one another. A tract home in a dense subdivision earns a tight band; a one-of-a-kind property in a thin market earns a wide one.
Reading the prediction interval
Because the model emits a full distribution, we can quote any coverage level. We surface two, chosen to match how a valuation is actually used:
- Most-likely range (50% interval).The central band the sale is as likely to land inside as outside — the CMA-style spread you would show a seller.
- Worst-case range (90% interval). The honest envelope for lending or risk: nine times out of ten the sale lands inside it. This is the
value_low/value_highpair in the response.
Why the two ranges aren't simple multiples of each other.Under the familiar bell curve the 90% band would be about 1.65× the 50% band. Because our distribution is heavy-tailed, the 90% band reaches into the fat tails and is closer to threetimes as wide — and that ratio varies from home to home. This is exactly why we quote intervals by coverage level(most-likely / worst-case) and never as a “2-sigma” multiple: a fixed sigma multiple would silently mean different coverage for different properties.
Interval widths are calibrated: we verify that a stated 90% interval really does contain the eventual sale price about 90% of the time, checking this separately across price levels and comparable-sale availability. As high-quality comps accumulate, the tails genuinely tighten — comp-rich homes carry markedly narrower bands than comp-sparse ones at the same price.
A wide band is information, not a defect: it is the model flagging a home where comparable evidence is thin and local expertise should carry more weight.
Off-market vs active-listing valuations
We run two complementary regimes sharing the same distributional machinery:
- Off-market. The default: value a home that is not currently for sale, with no asking price available. This covers the vast majority of the housing stock and is the regime all accuracy figures above describe.
- Active-listing.When a home is currently listed, its own asking price is a strong, leak-free signal — a seller and agent have already done a pricing analysis. Incorporating it roughly halves the error, and the model is explicitly held to a higher bar: it must beat simply echoing the asking price, which it does consistently.
Known limitations
- Luxury homes. Above ~$1M the model tends to underprice slightly, because prestige, finishes, and views are hard to observe from structured data. Bands widen accordingly and confidence is reported as lower.
- Very low price points.Below ~$200K, distressed sales, investor flips, and non-arm's-length transfers make outcomes intrinsically harder to price.
- Thin markets. Unique properties or areas with few recent comparable sales receive honestly wide intervals; the estimate should be treated as a starting anchor for a manual CMA.
- The estimate is a sale price. It predicts the likely transaction price, not a list price. A list price is typically a small markup on top, adjusted for strategy and market temperature.