How AI Predicts Property Values in Changing Markets

What Is an AI Property Valuation Model (AVM)?

An Automated Valuation Model (AVM) is an algorithmic system that estimates a property's market value by processing large datasets — sales records, property attributes, location signals, and economic indicators — without requiring an in-person inspection. The estimate is mathematical, not judgmental. AVMs update continuously as new data flows in, which is both their biggest advantage and a key source of error when the market moves faster than the training data.

The most widely known AVM is Zillow's Zestimate, but the technology is used across the industry. Major AVM providers include:

• Zillow Zestimate — consumer-facing, updated daily per Zillow's published methodology
• CoreLogic — widely used by lenders and servicers for mortgage underwriting
• ATTOM Data Solutions — deep county-level tax and deed records
• Redfin Estimate — MLS-connected, updates more frequently in markets where Redfin has listing access
• Black Knight (now ICE Mortgage Technology) — used heavily in secondary mortgage markets

How an AVM differs from a formal appraisal

A licensed appraisal involves a human professional who physically inspects the property, applies judgment to condition and upgrades, and signs off on a value for a specific lending purpose. An AVM has none of that — it cannot see the renovated kitchen, the deferred roof, or the flooding in the basement. Appraisals are legally accepted for mortgage underwriting; AVMs typically are not accepted as standalone replacements for purchase appraisals, though hybrid and desktop appraisals (which use AVM data as a starting point) are increasingly accepted for refinances.

Neither method is infallible. Appraisals can be influenced by appraiser bias and local market familiarity. AVMs are constrained by data quality and training recency.

What Data Does AI Use to Predict Property Values?

AVMs draw from a wide range of structured data sources. The quality and recency of each dataset directly determines how accurate the model's output will be for a given property type and location.

The main input categories are:

• Comparable sales (comps): Recent arm's-length transactions for similar properties in the same area — the single most powerful predictor in most AVM architectures
• Property characteristics: Square footage, bedroom/bathroom count, lot size, year built, garage, pool, property type (SFR vs. condo vs. multi-family)
• Tax assessment records: County assessor data provides a floor estimate and tracks improvements and parcel boundaries
• Neighborhood and location data: School district ratings, walkability scores, proximity to amenities, census tract demographics
• Economic indicators: Interest rate environment, local employment rates, population growth, permit activity
• Government indexes: The FHFA House Price Index (HPI) tracks quarterly appreciation by metro area and is freely available from fhfa.gov — a widely used benchmark for estimating home-value drift between sale dates
• MLS listing history: Days on market, price reductions, relisting counts — behavioral signals about how motivated sellers are in a given area

How Machine Learning Models Work for Property Valuation

Modern AVMs are not simple regression formulas. They use multiple machine learning techniques, often in combination, to extract patterns from millions of data points and generalize to properties where no direct comp exists.

Hedonic regression (the baseline)

Hedonic pricing models treat a home as a bundle of individual attributes — each bedroom adds X dollars, each additional bathroom adds Y dollars, proximity to a school adds Z dollars — and sum them to a total value. This approach is interpretable and still used as a sanity-check layer in many production AVMs, but it assumes linear relationships that don't hold in practice (the 6th bedroom adds far less value than the 3rd).

Gradient boosting: the workhorse of modern AVMs

Gradient boosting algorithms — most commonly XGBoost and LightGBM — have become the dominant ML technique for structured tabular data, including property valuation. The algorithm builds hundreds of shallow decision trees sequentially, with each tree correcting the errors of the previous one. Key advantages for real estate:

• Handles non-linear relationships (e.g., diminishing returns on bedroom count)
• Works well with mixed data types (numeric, categorical, geographic)
• Naturally handles missing values — critical when some comps lack specific fields
• Feature importance outputs help engineers understand which inputs drive the most variance

Neural networks for spatial pattern detection

Deep learning — particularly convolutional neural networks (CNNs) and graph neural networks (GNNs) — is increasingly used for spatial pattern recognition in property valuation. CNNs can extract value signals from satellite imagery (pool detection, roof condition, landscaping quality). GNNs can model the "neighborhood effect" — a property's value influenced by its spatial graph of nearby properties, not just a simple radius buffer.

Neural networks generally require more training data than gradient boosting and are harder to interpret ("black box"), which makes them less common as the primary model in regulated mortgage applications.

Ensemble models: combining approaches

Production AVMs at scale typically use ensemble approaches — stacking or blending the outputs of multiple models (a gradient boosted model, a hedonic baseline, and a spatial neural network) with a meta-learner that weights each model's contribution based on local data density. In dense urban markets with abundant comps, the gradient booster dominates. In sparse rural markets, the hedonic baseline gets more weight because there isn't enough sales data for the tree models to generalize reliably.

Feature engineering: recency weighting and distance decay

Raw comps are not treated equally. Well-designed AVMs apply:

• Recency weighting: A sale from 6 months ago counts less than one from last week in a moving market
• Distance decay: Comps within 0.25 miles carry more weight than comps 2 miles away, with the decay rate calibrated by market density
• Similarity scoring: A 3/2 SFR comp is weighted higher for a 3/2 SFR subject than a 5/4 estate comp
• Market-condition adjustments: Applying FHFA HPI or local index factors to "time-adjust" older comps to current market conditions

Where AI Property Value Predictions Are Least Accurate

Understanding AVM blind spots is as important as understanding what they do well. AVMs perform worst in four specific contexts — and agents need to know when to distrust the number.

Rural markets with sparse comps

All AVM accuracy depends on having enough nearby, recent, comparable sales. In rural markets where the nearest comparable sale may be 10 miles away and 18 months old, the model is essentially extrapolating rather than interpolating. Confidence intervals (when published by AVM providers) are significantly wider in low-density markets. If a provider doesn't publish a confidence score, assume rural estimates carry substantial uncertainty.

Unique and luxury properties

Luxury and one-of-a-kind properties — custom architecture, exceptional views, private airstrips, historic designations — have no true comps by definition. The model falls back to generic square-footage pricing, which systematically undervalues exceptional properties and overvalues those with expensive-but-illiquid features. Zillow has historically acknowledged wider error bands for high-value properties in its published Zestimate methodology documentation.

Post-disaster and rapid-change markets

When a wildfire, hurricane, or flood event reshapes local desirability, AVMs trained on pre-event pricing will be wrong — sometimes drastically. The model has no mechanism to instantly reprice based on news events; it must wait for new arm's-length sales to enter the training data. Similarly, in markets experiencing sudden demand shocks (a major employer announcement, a pandemic-driven relocation wave), AVMs systematically lag actual market prices until enough transactions close to recalibrate.

New construction before sufficient sales data exists

A new subdivision with only a handful of closed transactions gives AVMs almost nothing to work with. The model may anchor on the land value alone or rely heavily on permits and assessor records, both of which significantly understate market value for desirable new builds. AVM accuracy in new subdivisions typically improves substantially after 20–30 arm's-length sales close in the development.

How Market Conditions Affect AI Predictions

AVMs are trained on historical data, which means they describe the market as it was, not necessarily as it is. When macroeconomic conditions shift rapidly, the gap between AVM output and true market value can widen significantly.

Rising-rate environments

When interest rates rise sharply, buyer purchasing power drops. Demand compresses, but sellers often resist price cuts initially — creating a standoff that reduces transaction volume. With fewer closed sales to train on, AVM recency weighting has less fresh data to pull from. During the 2022–2023 rate spike cycle, some AVM outputs continued reflecting 2021 peak pricing long after active-listing prices had softened, because the settled comps still showed elevated values from contracts signed months earlier.

Inventory shocks and pandemic-era distortion

The 2020–2022 housing market produced an unusual clustering of transactions at unusually elevated price points. AVMs trained heavily on this data initially overestimated values in markets that subsequently cooled. This is a textbook example of "distribution shift" — the statistical distribution of the training data no longer matches the distribution of the current market. Engineers call this model drift, and it's one of the core operational challenges of maintaining a production AVM.

Recalibration windows

Responsible AVM providers monitor model performance in production and trigger recalibration when prediction error metrics exceed defined thresholds. Recalibration involves retraining (or re-weighting) models on more recent transaction data. Zillow and CoreLogic both publish methodology documentation that describes their update cadences, though the specific trigger thresholds are proprietary. Agents should check the "last updated" date on any AVM estimate they use — a 90-day-old estimate in a fast-moving market may be materially stale.

How Follow Up Ace Uses Public Data to Estimate Home Equity for Sellers

Follow Up Ace's Seller Radar feature applies a subset of AVM principles to help agents identify which contacts in their Follow Up Boss database are most likely to be ready to sell — without requiring agents to run manual comps on their entire contact list.

Seller Radar runs a nightly account-wide scan and computes an Ace Seller Score for each contact. The score is built from five verified inputs (weights from chat-app/utils/sellerIntentEngine.js):

• Tenure (35 pts): How long the contact has owned the property — longer tenure generally correlates with higher motivation to sell
• Estimated equity (30 pts): Modeled using a conservative mortgage amortization estimate adjusted by FHFA HPI appreciation data — labeled "(est.)" and not appraisal-grade
• FHFA appreciation index (15 pts): Area-level price appreciation from the federal FHFA House Price Index, one of nine federal datasets Ace ingests
• Area-risk signals (8 pts): Location-level risk factors drawn from the 30+ third-party data sources Ace connects to, including county parcel/assessor records across 25+ states via ArcGIS/Socrata
• CRM seller tags (12 pts): Explicit signals from the agent's own Follow Up Boss data — tags, notes, or stage assignments indicating seller intent

The nine federal datasets Ace uses (via connectors in chat-app/utils/publicData/connectors/) include: FHFA HPI, Census ACS, FEMA flood zones, openFEMA disaster data, USFS wildfire risk, USGS seismic data, CFPB delinquency data, HUD housing data, and Census geocoder. These feed into the 30+ third-party data source total.

A few important limitations to understand about the Ace Seller Score:

• Home equity is estimated (est.) using a modeled mortgage calculation — Ace does not access lien data or observed payoff amounts, and the figure should never be treated as appraisal-grade precision
• The Ace Seller Score surfaces in the Seller Radar dashboard inside the Follow Up Ace embed — it is not currently synced to a Follow Up Boss custom field (that write-back is on the product roadmap)
• Suggestions from Seller Radar are reviewed by the agent before any outreach — nothing is auto-sent

For agents who want to understand the broader Ace Trove capabilities powering these data integrations, or who want to explore agentic workflows that surface these insights in context, the product documentation covers both in detail.

What Real Estate Agents Should Know About Trusting AI Valuations

For agents advising clients, AVMs are a useful starting point — not a conclusion. Here is a practical framework for using them responsibly:

1. Use AVM estimates to frame the conversation, not close it. An AVM gives you a data-backed ballpark to anchor the CMA discussion. Clients who have already looked up a Zestimate will be more receptive when you acknowledge the number and then explain the variance.
2. Always cross-reference with local MLS comps. Pull your own comps within the past 90 days, ideally within 0.5 miles and the same property type, adjusted for condition. The AVM cannot see condition; you can.
3. Check when the estimate was last updated. A 60–90 day lag in a rising or falling market can mean a material price difference. Look for the "data as of" or "last updated" timestamp before citing the number.
4. Understand the confidence interval if it's published. Zillow publishes Zestimate accuracy statistics by market on their website. A market where the median error rate is 8% represents a different risk profile than one where it's 2%.
5. Flag the blind spots to clients. If the property is unique, rural, or recently renovated, tell clients explicitly that the AVM estimate is less reliable than it would be for a standard suburban resale property.
6. Treat AI equity estimates as directional, not definitive. Whether from a commercial AVM or a tool like Ace Seller Radar, equity estimates based on modeled data should prompt a conversation — not replace a formal payoff request or CMA.

For deeper context on how agents can build data-driven client conversations, see our guides library and the Ace blog.

Frequently Asked Questions

How accurate is the Zillow Zestimate?

Zillow publishes accuracy statistics on their website broken down by market. Accuracy varies significantly by region and property type. For on-market homes (where Zillow has listing data), accuracy is typically higher because the listing price itself is a strong signal. Off-market accuracy is lower and varies widely. Always check Zillow's published methodology for the specific market you are working in rather than relying on a national average.

Can AI replace a real estate appraisal?

Not for most purchase transactions. Licensed appraisals remain required by lenders for purchase mortgages in the vast majority of cases because they provide legal liability coverage that AVMs do not. Hybrid and desktop appraisal waivers (where the lender accepts an AVM supplemented with a property data collector visit) are increasingly available for refinances and certain low-risk loan products, but the standards are set by the GSEs (Fannie Mae / Freddie Mac) and vary by loan type and borrower profile.

What is the FHFA House Price Index and why does it matter?

The FHFA House Price Index (HPI) is published by the Federal Housing Finance Agency and tracks quarterly home price appreciation across U.S. metro areas using repeat-sale transactions. It is freely available at fhfa.gov. Because it measures the same properties over time (controlling for property mix), it is considered one of the most reliable measures of true appreciation. AVMs and tools like Ace Seller Radar use FHFA HPI to time-adjust older comps and to model how much a property purchased several years ago has likely appreciated — even without a recent sale.

How often do AVMs update?

Update frequency varies by provider. Zillow updates Zestimates daily for on-market homes and regularly for off-market properties, per their published methodology. CoreLogic and other institutional AVM providers typically refresh at varying cadences depending on the product and data feed. The key factor is not just update frequency but recalibration frequency — how often the underlying model is retrained on new transactions. These are often different numbers and may not be disclosed publicly.

Should agents share AVM estimates with clients?

Yes — transparently. Clients will find these estimates on their own, so the more effective approach is to share them proactively and explain the methodology and limitations. This builds trust and positions you as the expert who can contextualize the data, rather than someone who appears to be hiding an inconvenient number. Frame AVM estimates as "one data point" alongside your CMA, your local market knowledge, and any condition factors the AVM cannot see.