Automated Valuations and Why Claremont Confuses Them

Every homeowner has looked one up. Type an address into a portal and a value appears, confident and specific, updated whenever the model runs again. These automated valuation models are genuinely useful tools used for the right purpose, and they are the source of an enormous amount of misplaced certainty when used for the wrong one.

What an automated model actually does

An automated valuation model is a statistical estimate. It ingests public records, recorded sale prices, tax data, listing history where available, and property characteristics, then applies a model trained on large volumes of transactions to predict what a given property would sell for.

Nobody looks at the house. Nobody verifies the record. Nobody adjusts for the condition of the kitchen or the fact that the addition was finished last spring. The model works from data, and the data is public record data with the strengths and gaps that implies.

This is a fundamentally different exercise from an appraisal, in which a licensed professional selects comparable sales by judgement, adjusts for specific differences, reconciles among recognised approaches to value, and signs a certification carrying professional accountability. The two are not competing versions of the same thing.

Where the models are strong

Credit is due. On a large, uniform housing stock with high turnover, automated models perform well. A tract of similar homes trading regularly gives a model exactly what it needs: many recent transactions of properties that genuinely resemble each other, so a statistical inference across them is well-supported.

They are also useful at scale, for portfolio monitoring, for market-level trend observation, and as a rough starting point for a homeowner idly wondering where things stand. None of those uses depends on any individual estimate being precisely right.

Why they struggle here specifically

Claremont supplies almost every condition that degrades an automated estimate.

THIN TRANSACTION DATA. Long tenures mean fewer recent nearby sales for the model to learn from. A model with sparse local data leans on data from further away or further back, which is exactly where its accuracy falls off.

HETEROGENEITY AT SHORT RANGE. Housing eras change block to block here. A model that treats nearby properties as broadly similar will average across houses that no buyer would consider substitutes for one another.

PUBLIC RECORD DRIFT. Older housing that has been altered across decades frequently carries an assessor record that no longer describes the house. If the recorded square footage predates a permitted addition, the model is valuing a property that does not exist. That drift is common enough here to be assumed rather than checked for, and how a real measurement is produced is covered in the square footage article.

CONDITION IS INVISIBLE. Two identical floor plans, one meticulously maintained and one with three decades of deferred maintenance, look the same to a model. The gap between them is the sort of thing an appraiser records in a single afternoon.

SITE FEATURES DO NOT ENCODE. A view, a slope, a corner lot, proximity to a busy arterial, an unusually deep parcel - value-relevant, and poorly captured in structured data. Hillside property is the sharpest example, and the view-home article explains why even human appraisers find it demanding.

Read the confidence range, not just the number

Most reputable providers publish a confidence indicator or a value range alongside the headline estimate. Almost nobody looks at it, and it is the most informative part of the display. A wide range is the model telling you honestly that it does not have enough comparable local data to be precise. On distinctive property in a low-turnover town, wide ranges are the normal case rather than the exception.

Providers also differ. Running the same address through several services and getting materially different answers is not evidence that one is broken; it is evidence that different models weight different inputs where the data is thin.

What to use them for, and what not to

USE THEM as a rough orientation, as a way to watch broad direction over time, and as a prompt to have a real conversation.

DO NOT USE THEM to set a list price, to decide whether an offer is fair, to conclude that an appraisal was wrong, or as evidence in any formal process. A reconsideration of value request built on an online estimate is a request built on nothing, because the model's inputs cannot be examined, verified, or defended - the material that actually works is set out in the reconsideration article.

It is worth adding that lenders do use automated valuations in their own processes, in controlled ways with their own accuracy testing and thresholds. That is a supervised institutional use with policy around it, and it is not the same thing as a consumer reading a portal estimate.

The document that actually answers the question

For a decision that matters - a list price, an offer, an estate division, a refinance strategy - the useful inputs are a comparative market analysis prepared by someone who watches this market, or a formal appraisal by a licensed professional, depending on the purpose. What each one is for is the subject of the appraisal versus CMA article.

Anthony prepares a comparative market analysis grounded in the actual comparable sales and the actual condition of the property, and coordinates independent, state-licensed appraisers where a formal appraisal is required; he does not perform appraisals and does not influence them. Anthony Grynchal has been licensed in California since November 2009.

For the full cluster, start at the appraisal guide, and read the comparable-sales article next if you want to understand why the local data is thin in the first place.