The Training Data Is Being Withdrawn

Every automated valuation in residential real estate reads a public record. That record is currently shrinking, and it is shrinking in a specific place: the fields that describe how the market responded to a price.

A growing share of homes now sell as private exclusives or coming-soon listings rather than on the open market. On those sales, two things that used to be visible are not: how long the property was for sale, and whether the price was ever reduced. The house still trades. The account of how it traded does not enter the file.

The two fields carry more than they appear to

Days on market and price-cut history are not descriptive colour. They are the only fields in the record that report a verdict rather than an attribute. Square footage says what a house is. Time on market says what the market thought of the price it was offered at.

A model that reads both is doing something meaningfully different from a model that reads only the first. With price history, it can distinguish a home that cleared quickly at asking from one that sat 90 days and conceded twice, and it can learn what that distinction means for the next property. Without it, both sales enter the training set as a price and a set of attributes, and the difference between them is unrecoverable. The gap sits alongside the other structural omissions in what the public record gets wrong, with the difference that this one is widening rather than holding steady.

The confidence does not move

Here is the part that matters for anyone building or evaluating a valuation model. As the share of off-market sales rises, the inputs available to the model degrade continuously and quietly. There is no threshold event, no version where the data becomes visibly insufficient. Coverage thins by a few points a year.

Nothing in a typical automated estimate is designed to report that. The output format is fixed: a number, sometimes a range, usually a confidence label calculated from comparable agreement rather than from comparable completeness. A model can be highly confident because its comps agree with each other while every one of those comps is missing the field that would have shown the price was wrong. That is the failure mode set out in what confidence should mean in a valuation tool, arriving now through the data supply rather than through the method.

So the trajectory is a system that becomes less informed and no less certain, and reports the second while concealing the first.

A second-order effect on the industry that built the models

The same withdrawal that thins the models also thins everyone else. An agent pricing a home in a market with heavy off-market volume is working from a comp set where some of the sales have no story attached. They cannot tell which of their comps was a clean sale and which was a rescued one, and neither can anyone else.

That is a different problem from the usual complaint about data quality. This is not bad data. It is absent data, withdrawn deliberately, described as privacy, and it removes the same evidence from the professional and the algorithm at once. The difference is that the professional can go and ask. There is still a person who knows what happened on that listing, and they can be called. No model has that option.

How do off-market sales affect automated home valuations?

They remove days on market and price-reduction history from the record the model reads, while leaving the sale price and property attributes intact. The model can still price the home against the sale, but it cannot see whether the market accepted or resisted the price along the way. As the share of off-market sales grows, that blind spot grows with it.

Does a confidence score reflect missing data?

Usually not. Confidence in most automated valuations is derived from how closely the selected comparables agree with one another, not from how complete the record behind each comparable is. A set of comps that all lack price history can agree tightly and produce a high confidence figure, because the measure was never designed to report what is absent.

Why do days on market and price cuts matter for pricing?

They are the market's response to a price rather than a description of a property. A home that sold in a week at asking and one that sat three months and reduced twice are reporting opposite things about how their prices were set. Without those fields, a listing the market rejected and one it never had the chance to reject look identical in the file.

Less information, same voice

The industry conversation about valuation models is mostly about method: better comp selection, better adjustment, better modelling of local effects. The input question is getting less attention and it is the one moving fastest. A method improvement of a few percent means little against a data source that is being narrowed by design.

This is Context Blindness™ at the level of the whole market rather than one house. It is not that a model missed a fact about a specific property. It is that the facts are being withdrawn from the file, across the board, while the estimates continue printing in exactly the tone they used when the file was complete. A reasoned valuation, the approach CMAflow builds, treats the completeness of the evidence as part of the output rather than a hidden assumption, and widens where the record thins. The market will keep getting quieter. The screens will not sound any less sure.

---

This article is general information and analysis, not financial, lending, or appraisal advice. Verify any home value with a licensed professional before acting.

The Independent Agent
Substack | Spotify | CMAflow FAQ | YouTube | Free CMA | Home valuation | Insights | Blog