Auction results need transaction context before machines learn from them

A transaction dataset records observed exchanges under particular sale and participation conditions. It is not a direct inventory of the amounts every owner could realize at any time.

If a model learns sale prices while dropping the conditions that helped produce them, it can present a prediction of observed transactions as an unconditional statement of asset value. The error is in the target being claimed, even where the recorded prices are accurate.

Completed-sale data conditions on completion

Consider two constructed datasets. Both contain the same successful auction sales. One also records offered lots that did not sell, their reserves, and the offering dates. The other omits them.

The second dataset can describe prices among its recorded completed sales. It cannot estimate a lot’s probability of selling from those records alone, because it lacks the offered-but-unsold outcomes and the relevant denominator. A high predicted price conditional on sale is not a high probability of obtaining that price.

This is a conditional limitation of that dataset design. It does not establish that every public auction database omits unsold lots or quantify a universal bias in published results.

Transaction terms change the meaning of price

Under the auction terms, an irrevocable bid establishes a purchase commitment if higher bidding does not prevail. A permitted fee can be netted against the successful irrevocable bidder’s purchase price. Interested parties also have economic relationships to a lot that unaffiliated bidders do not share.

Those conditions matter to the question being predicted. The displayed amount is not automatically the successful buyer’s net economic outlay, and a supported sale is not a measurement of the counterfactual outcome without that support.

Price basis also needs to distinguish the reported sale amount from seller proceeds. A model cannot infer missing transaction charges simply by seeing more headline prices.

Public information need not be fully reflected in observed prices

US housing research comparing floodplain designations with transactions found that mapped flood information was not fully capitalized at the level implied by the study’s valuation model. Commercial buyers and transactions in stronger disclosure settings showed different price associations.

Those comparisons do not isolate information as the only cause. They do establish that publicly available information and its observed incorporation into prices are different objects of analysis. The housing results do not supply a correction factor for art or wine.

Specify the prediction before fitting the data

A conditional sale-price prediction, a sale-probability estimate, and net proceeds under a deadline require different outcomes and covariates. The US fair-market-value assumption of no compulsion also differs from a transaction forced by distress.

An interpretable model states its price basis, covered transactions, timing, missing outcomes, and relevant sale conditions. Its output can then answer the question its data supports. Removing those qualifiers does not broaden the model’s evidence; it broadens the unsupported claim attached to its number.