# Automated valuation models and the appraisal revolution
In November 2021, Zillow shut down Zillow Offers, its iBuying (instant buying) business, and announced it would wind down operations and cut roughly a quarter of its workforce. The company had been using its own valuation engine to buy homes at scale, then resell them. When the model's price predictions drifted from reality in a fast moving market, Zillow ended up holding thousands of homes it had overpaid for. The unit lost hundreds of millions of dollars.
The lesson was not that automated valuation is useless. It is that a valuation model is only as good as its confidence intervals, and Zillow bet real capital on numbers that were less certain than they looked.
An AVM (Automated Valuation Model) estimates a property's market value using data and statistics instead of a human walking through the house. Lenders, insurers, tax assessors, and portals like Zillow and Redfin all use them.
The famous consumer example is Zillow's Zestimate. But AVMs also sit quietly inside mortgage underwriting, portfolio risk management, and property tax appeals.
Every AVM combines three ingredients:
Let's take each apart.
A human appraiser picks three to six recent sales nearby, adjusts for differences ("this one has an extra bathroom, add value"), and lands on a number. This is the sales comparison approach, and it is the backbone of residential appraisal.
An AVM does the same thing at scale. Instead of six comps, it might weigh hundreds, weighting each by similarity and recency.
The weakness is obvious: comps assume the recent past predicts the present. In a stable market that holds. In a turning market (rates spiking, a local employer leaving), yesterday's sales mislead. This is precisely where Zillow's iBuying model struggled: prices moved faster than the comps could refresh.
A hedonic model treats a house as a bundle of features and estimates the price contribution of each one. The idea comes from economics: you are not buying "a house," you are buying square footage plus a garage plus a good school zone.
A simplified hedonic regression looks like this:
import statsmodels.formula.api as smf
# price as a function of measurable features
model = smf.ols(
"price ~ sqft + bedrooms + bathrooms + lot_size + age + C(zip_code)",
data=sales
).fit()
# the coefficient on sqft = estimated dollars per additional square foot
print(model.params["sqft"])The coefficient on sqft tells you how much an extra square foot is worth, holding other features constant. Modern AVMs replace this simple regression with gradient boosting or neural networks that capture nonlinear effects (the tenth bathroom adds far less than the second).
The catch: models only see features that are in the data. A gut renovated kitchen and a dated one may have identical square footage. Condition, light, layout, and street noise are hard to digitize, and they can swing value materially. Assessors call this the unobserved quality problem, and it is a permanent gap between the model and the market.
Two identical houses can differ 30 percent in value across a street that divides school districts. Location is not one variable; it is thousands of overlapping ones.
Good AVMs ingest:
The technical challenge is that space is not linear. A model that treats latitude and longitude as ordinary numbers will fail. Techniques like geographically weighted regression or spatial features (neighborhood embeddings, distance decay functions) exist specifically to handle this.
For a clear, free primer on how spatial autocorrelation works, the Geographic Data Science book is an excellent open resource.
🎬 [VIDEO: "How Zillow's Zestimate Works" — youtube.com — a plain-language walkthrough of the data and methods behind consumer AVMs]
Here is the part most people skip. An AVM does not just output a price. It should output a confidence interval: a range and a probability, for example "we estimate $500,000, and we are 80 percent confident the true value is between $460,000 and $540,000."
Zillow publishes a median error metric for the Zestimate on-market and off-market homes. That published error is an average. Averages hide the cases that hurt you.
Confidence intervals break down in predictable situations:
Thin markets. Rural areas or unique luxury homes have few comps. Less data, wider (and less reliable) intervals. A one of a kind waterfront estate has no true comparable.
Regime change. When the market turns, the model's training data describes a world that no longer exists. The interval says "confident" because it was trained on a stable era.
Selection bias. iBuyers face adverse selection: sellers are more likely to accept an instant offer when the AVM has overpriced their home, and more likely to reject it when the offer is too low. So the homes you actually buy skew toward your model's overvaluations. This is a structural trap, not a tuning problem, and it is widely cited as a factor in iBuying losses.
Renovation and condition drift. The model assumes average condition. Buy a home that is worse than average, and you overpaid, systematically.
The strategic point: a point estimate invites overconfidence. A well calibrated interval tells you when to send a human.
Knowledge check
1. What was the core lesson of Zillow's failed iBuying experiment for practitioners deploying automated valuation models?
2. An AVM's comps engine is best described as an automated version of which traditional appraisal method?
3. How does an AVM's use of comps typically differ from a human appraiser's approach?
4. Select ALL correct answers about the three core ingredients that every AVM combines.
Select all the correct answers.
5. Select ALL correct answers about where AVMs are used according to the lesson.
Select all the correct answers.
AVMs are reshaping the appraisal industry, but "revolution" does not mean appraisers vanish.
Regulatory reality. In the United States, federal rules require appraisals for many mortgage transactions, though appraisal waivers (approved by Fannie Mae and Freddie Mac and often powered by AVMs) have expanded for lower risk loans. In 2024, U.S. federal agencies finalized a rule requiring quality control standards for AVMs used in mortgage decisions, including safeguards against discrimination. So AVMs now operate inside a compliance framework, not a free for all.
Fair lending risk. AVMs learn from historical data, and historical housing data reflects historical discrimination (redlining, biased appraisals). A model that ingests this can reproduce it. This is a live regulatory concern, and the 2024 AVM rule explicitly targets it. For real estate professionals, "the algorithm did it" is not a legal defense.
The emerging workflow. The practical model in 2026 looks like this:
Appraisers who understand the models will supervise them. Those who treat their craft as immune to automation face the most pressure.
Zillow's failure was not a data science failure in isolation. It was a business failure of trusting a point estimate with capital, ignoring the interval, and underpricing the structural bias of buying homes based on your own model.
The transferable lesson for anyone in real estate: when a vendor sells you an AVM, ask two questions. How wide is the confidence interval on my kind of property? And what happens to that interval when the market moves?