When two AVMs disagree: a lender's investigation guide

A second valuation model is most useful when disagreement leads to a better investigation. Simply averaging two estimates can hide the reason they differ.

A lender receives two automated estimates for the same property: £300,000 and £360,000. Both arrive quickly and look plausible. The easy response is to choose the midpoint, but £330,000 is still an estimate without an explanation.

The disagreement may reveal different assumptions, evidence or model behaviour. Understanding which of those applies is more useful than making the discrepancy disappear.

Check that both outputs answer the same question

Confirm the property identity first. A parent building, an individual flat and a similarly addressed unit can produce very different values. Review the match before debating the modelling.

Then compare the valuation date and the meaning of each output. A predicted listing price, an estimated transaction value and an assessed value under specified assumptions are not automatically interchangeable. Chimnie's documentation distinguishes its estimated listing-sale field from other property information; buyers should preserve the stated definition when making comparisons.

Check the property description supplied to each model. Floor area, tenure, bedrooms, condition and recent alterations may differ. Also identify any confidence or uncertainty information that accompanies the estimate, and the population on which that information was validated.

Investigate the evidence behind the gap

Return to the fictional £300,000 and £360,000 estimates. Suppose one model uses an older floor area and a previous sale, while another incorporates evidence of additional accommodation. The relevant next step is to establish the current property configuration and whether the additional space is appropriately reflected.

Alternatively, both models may describe the same property but rely on different comparables or respond differently to a thin local market. Agreement on the inputs does not guarantee agreement on the estimate.

Completed sales are valuable evidence, but their availability has timing constraints. HM Land Registry's Price Paid Data guidance explains the dataset's scope, which concerns sales in England and Wales submitted for registration. A recently completed transaction may not yet be present in the data available to a model.

Record what has been checked and what remains unresolved. Where appropriate under the lender's policy, obtain further evidence or a professional valuation. A second automated result should help target that work.

Validate the disagreement rule itself

If a lender plans to refer cases when two models diverge, that rule needs evaluation. A threshold that catches useful discrepancies may also generate substantial review volume in unusual or poorly evidenced properties.

Test how the rule performs across the intended population and against suitable outcomes. Examine the cases where both models agree but are wrong, as well as those where they disagree. Two models sharing important data sources may share errors, so agreement is not independent confirmation.

Track the cost and usefulness of the resulting referrals. A good second-source strategy might improve the selection of cases for investigation, provide an additional benchmark or expose inconsistent property records. Its value need not depend on replacing the lender's primary model.

Chimnie can be considered as part of that evidence-gathering process, with field definitions and the proposed role agreed at the outset. The decision remains tied to the lender's valuation policy and the evidence available for the individual property.

Speak to Chimnie about a defined second-source evaluation, including which disagreements would be useful for your team to resolve