The model in one equation
A simple hedonic regression looks like this:
Price = β0 + β1·sqft + β2·beds + β3·baths + β4·lot
+ β5·age + β6·location_score + ε
Each β (beta) is the marginal price of one unit of that attribute, holding everything else constant. Fit on a few hundred recent sales in a market, the model produces estimates like:
- β1 (sqft): $185 / ft²
- β2 (beds): $9,500 / bedroom (after controlling for sqft)
- β3 (baths): $14,000 / full bath
- β4 (lot): $4 / ft² of lot
- β5 (age): -$420 / year of age
Plug in the target's attributes, sum, and you have a value estimate. Real production AVMs use 50–200 features and non-linear extensions (splines, interactions, Gradient-Boosted Trees), but the spine is hedonic.
Why it works for real estate
Real estate is the canonical hedonic-pricing problem. Sherwin Rosen's 1974 paper "Hedonic Prices and Implicit Markets" formalised the idea: a home is a bundle of characteristics, not a single good, and the market reveals the implicit price of each characteristic through sales data. Three reasons it dominates modern AVMs:
- Scales to large datasets. Sales-comparison logic needs you to hand-pick comps; hedonic models fit on thousands of sales at once and extract the average attribute price.
- Handles atypical homes. When the target is a 5-bed home in a neighbourhood of 3-bed homes, sales-comparison runs out of comparables. Hedonic regression extrapolates using the bedroom coefficient.
- Produces confidence intervals. The regression's residual variance gives you a defensible error band. AVMs that don't show a confidence band are usually using simpler methods that can't.
Hedonic regression vs sales-comparison approach
The two approaches answer the same question different ways:
| Dimension | Hedonic regression | Sales-comparison |
|---|---|---|
| Data scale | 100–10,000 sales | 3–8 hand-picked comps |
| Adjustment logic | Coefficient per attribute | Line-item per comp |
| Best for | Typical homes in active markets | Atypical homes, expert review |
| Transparency | Black box without methodology page | Every adjustment visible |
| USPAP-compliant | No (statistical) | Yes (when done by appraiser) |
| Used by | Zillow, Redfin, HouseCanary, Twellie | Licensed appraisers, USPAP |
A modern AVM typically blends both: a hedonic / GBT model for the central tendency, plus an explicit comparable-sales view in the report. The line-by-line adjustments are what makes the math auditable — see comp adjustment factors explained for the full breakdown.
Where hedonic models break
Three failure modes worth knowing:
- Multicollinearity. Sqft and bed count are correlated; the model can't cleanly separate the two coefficients. Solution: regularisation (Ridge / Lasso) or feature selection.
- Non-linearity. Adding a 4th bedroom to a 1,200 ft² house doesn't add the same value as adding a 4th bedroom to a 3,000 ft² house. Linear hedonic regression misses this; GBT and spline models capture it.
- Sparse markets. With under ~50 sales in a 12-month window, the regression coefficients are noisy. This is why rural and thin-market valuations have wider confidence bands.
The third one is structural — no model fixes a data problem. In sparse markets the appraiser's hand-picked comparables (with explicit adjustments) often beat the AVM. See our AVM vs appraisal vs Zestimate guide for when to use each.
How to spot hedonic-driven AVMs in the wild
When you read an AVM's methodology page, the giveaway is language about "feature coefficients," "hedonic price index," or "attribute-level pricing." All three of these mean hedonic regression at the core. Zillow's Zestimate methodology, the S&P CoreLogic Case-Shiller index, and most academic real-estate price indices are hedonic-based.
The Federal Housing Finance Agency (FHFA) House Price Index uses a repeat-sales method instead, which avoids hedonic estimation by tracking the same homes over time. Different tradeoffs: more robust to attribute estimation error, but blind to home improvements between sales.
What this means for your offer
You don't need to do the regression yourself. What you need to know:
- If the AVM's confidence band is wide (>10% of value), the local hedonic model is fitting a sparse or heterogeneous market. Trust the appraisal more.
- If two AVMs disagree by less than 5%, both their hedonic models are converging on similar coefficients — that's signal.
- If you're buying an atypical home (custom, unusual lot, rare configuration), the hedonic model is extrapolating; verify with hand-picked comps before offering.