Why Home Value Estimates Break on Fixer-Uppers

Automated home value estimates break on fixer-uppers because the single most important variable, condition, does not exist in the data these models are built on, and because the comparable sales they draw from are dominated by homes in better shape. The result is not random noise, it is a systematic bias in one direction: fixers come out valued too high. That directional bias is the practical problem. Random error can be absorbed by a margin of safety. A predictable upward bias on exactly the property type investors buy is a mechanism for overpaying, repeatedly, with software producing the confidence.

Why can't the model just see that the house needs work?

Because nothing in the underlying record describes condition. A county assessor file lists the address, the living area, the bed and bath counts, the year built, the lot size, and a property class. There is no field for roof age, no field for whether the kitchen has been touched since it was installed, no field for a foundation issue, no field for whether the house has been vacant for two years. The consequence is that every AVM has to assume something about condition, and the only defensible default assumption is roughly average for the area. Applied to an average house, that is correct. Applied to a fixer, it produces a number describing a house that does not exist: the subject property in condition it is not in.

Why does the comp pool make it worse?

This is the second half of the problem and the less obvious one. Many valuation pipelines filter distressed, as-is, and non arm's length sales out of the comparable pool. The reasoning is sound in the general case: an estate sale, a foreclosure auction, or a heavily discounted as-is transfer may not reflect open market value, and letting them into the comp set for a normal home would drag the estimate down incorrectly. But that same filter, applied to a fixer-upper, removes precisely the sales that describe what the subject is worth. The model is left comparing a house that needs sixty thousand dollars of work against a pool of homes that have already had that work done. The estimate that comes back is much closer to after repair value than to as-is value, and nothing in the output says so. Add the third effect, regression toward the mean, which pulls predictions on low priced properties upward toward the center of the training distribution, and all three biases point the same way.

Why doesn't the listing text solve it?

It helps, and it is far from sufficient. Listing descriptions are marketing documents. "Handyman special," "bring your vision," "sold as-is," and "priced accordingly" are useful signals when they appear, but agents write them inconsistently, they appear on properties in wildly different states of repair, and plenty of genuine fixers are described in entirely neutral language. There is no standard vocabulary and no requirement to disclose condition in the description at all. More fundamentally, text tells you an agent's framing, not the property's state. A house described as "well maintained" and a house described as "needs TLC" can be closer in actual condition than either description suggests.

What about photos?

Photos are the strongest condition signal in the public record, because they show the thing directly instead of describing it. Finish level, flooring, cabinetry, fixture age, visible damage, and general upkeep are all legible in a standard listing photo set in a way they are not legible in any structured field. This is why we estimate condition from listing photos and use that read to separate as-is value from after repair value. Two things are worth being clear about. First, it is an estimate and not an inspection: a photo set will not show a failing sewer line, a cracked slab, or knob and tube wiring behind a wall, and sellers do not photograph the parts they would rather you not see. It is a real improvement over assuming average condition. It is not a substitute for walking the property, and we would rather say that plainly than imply otherwise.

As-is value and ARV are different questions

The distinction matters more on a fixer than on any other property type, because on a fixer the gap between the two is the entire deal. After repair value is what the property sells for once the work is done and it is marketed normally. This is the question automated valuation handles best, because the comp set of renovated sales is dense and well documented. As-is value is what the property is worth right now, in current condition, to the pool of buyers who will actually bid on it in that condition. That pool is smaller, more sophisticated, and pricing in risk, timeline, and financing constraints as well as the raw cost of work. The gap between them is not simply the rehab budget. It also includes holding costs, the contractor's and investor's margin, the risk premium buyers apply to unknown condition, and the reality that fewer lenders will finance a property in poor condition, which narrows the bidder pool and the price. If a tool shows you an as-is value roughly equal to its after repair value on a property that visibly needs work, that is the clearest possible signal that the model has no condition input. Treat the as-is number as unusable.

What should I do when valuing a fixer-upper? 1. Use the automated number as ARV, not as as-is value. It is much better at the finished house than the current one. 2. Build the rehab budget separately and honestly. Scope from photos and a walkthrough, not from a per square foot rule of thumb. The rehab cost estimator gives a starting framework you can adjust. 3. Derive as-is rather than reading it off a screen. Start from ARV, subtract rehab, subtract holding and transaction costs, subtract the margin required for the risk. What remains is what the property supports. 4. Look at the comps you were given. If every one of them is renovated, the number describes renovated homes. Exclude them and see what happens. Comp sets you can adjust are the point. 5. Widen your range. On a property with real condition uncertainty, a single figure is false precision. Underwrite a range and buy against the pessimistic end. 6. Get eyes on it. No model, including ours, sees the foundation. Run the finished numbers with the free ARV calculator, and take the full deal through the deal analyzer so the rehab budget, the exit assumption, and the value estimate stay attached to each other.

Why we publish where this breaks

The failure mode described here is real, it is present in our model as well as everyone else's, and the only reason we can talk about it specifically is that we measure it. Every estimate is frozen when a property lists and graded against the actual closing price when the home sells, and the results, including the property types where performance is worst, are published at /accuracy. An accuracy page that only shows you the flattering segments is a brochure. The distressed and low priced segments are where the interesting information is, and they are the ones investors should read first.