Skip to main content
Back to timeline
arXivSource publication:

Auction reserve prices now come from listing text: across roughly 65,800 used-vehicle auctions, the net automatic-sale revenue gap reaches $165.8 million

Synopsis

Online auction houses must set reserve prices for highly differentiated products, while established approaches require managers to specify product attributes in advance. This work fine-tunes a pretrained text embedding model against the structural likelihood of auction bids, so the description representation targets the valuation distribution that determines the reserve, then estimates a generalized Pareto tail on the other half of the sample and derives the reserve in closed form. Across roughly 65,800 US online used-vehicle auctions, with seller values calibrated to match the sale rate, the model implies a counterfactual net automatic-sale revenue increase of $165.8 million, or 14.6% above the incumbent reserve policy.

Source-provided article image: Setting Optimal Reserve Prices using Fine-Tuned Language Models
arXiv

Interpretation

A reserve price can be issued listing by listing from the description and a seller value, without building a product taxonomy and without a bid history for the item being priced. Earlier routes either required managers to specify product attributes in advance or relied on repeated auctions of comparable goods or bidder-specific histories to infer the valuation distribution. Across roughly 65,800 US used-vehicle online auctions from 2014 to 2023, with seller values calibrated to match the incumbent 72.3% sale rate among reserve listings, the model's median reserve is $20,686, with an interquartile range of $12,978 to $33,947.

The model condenses each listing description into a low-dimensional representation, and that representation sets the tail shape of the listing's valuation distribution. The training signal is the structural likelihood of observed bids rather than a forecast of the final price, so representation learning is directed at the valuation distribution that determines the reserve. Stage I fine-tunes Qwen3-Embedding-0.6B with a low-rank adapter on one random half of the sample and freezes it; Stage II maps the embedding to the shape and scale of a generalized Pareto tail on the other half, the inverse hazard is linear, and the reserve follows in closed form.

Against the incumbent policy, the model's reserves move two groups of auctions in opposite directions. Under the incumbent policy, 4,194 auctions had a final bid above the model's reserve but below the seller's reserve, and another 4,194 sold although the final bid fell below the model's reserve. The first group represents $266.7 million in observed final bids convertible to automatic sales, the second places $101.0 million in realized revenue at risk, and the net difference of $165.8 million equals 14.6% of the $1.134 billion in observed revenue under the incumbent policy.

Perspective

The findings cover roughly 65,800 auctions held from 2014 to 2023 on one US online used-vehicle auction platform, with seller value treated as a policy input calibrated to match the incumbent sale rate among reserve listings; both the reserve level and its intervals are conditional on that choice. The revenue comparison covers only automatic sales at auction close, and outcomes of private negotiations after an unmet reserve are unrecorded, so it does not identify effects on total transaction revenue or seller payoff.

Count-free identification requires symmetric independent private values and a valuation distribution that does not vary with market size after conditioning on the description representation; unobserved listing quality or common information could shift valuations and participation together. The retention rule selects terminal bids most likely to reveal valuations by timing and threshold crossing, and if that selection correlates with unobserved valuations the estimated valuation tail may be biased. The base embedding model's pretraining corpus cannot be fully verified, so the sample split cannot rule out prior exposure to inference-half observations.