TL;DR
This paper presents a novel framework for setting optimal auction reserve prices by fine-tuning language models directly on the structural likelihood of historical bids. By extracting demand signals from unstructured listing text, the authors demonstrate a potential 14.6% revenue increase ($165.8 million) in a large-scale used-car marketplace compared to traditional human-led pricing policies.
What Problem Does This Paper Address?
In online marketplaces for highly differentiated goods (like vintage cars or antiques), setting the "reserve price"—the minimum price a seller will accept—is a high-stakes challenge. If the reserve is too high, the item doesn't sell; if it's too low, the seller leaves money on the table.
Traditional economic models for reserve pricing usually require a "product taxonomy"—a fixed set of attributes like year, make, and mileage. However, much of an item's value is hidden in unstructured text: a "meticulously maintained" engine or "minor surface rust" can swing valuations by thousands of dollars. Furthermore, standard structural models often require knowing the total number of potential bidders ($N$), which is rarely observable in the real world. The authors aim to bridge the gap between deep learning's ability to process text and structural econometrics' ability to model optimal decisions.
Methods
The authors propose a two-stage framework that integrates a Large Language Model (LLM) into a structural economic model:
- Stage I (Structural Fine-Tuning): They take a pretrained text embedding model (Qwen3-Embedding-0.6B) and fine-tune it on one half of their dataset. Crucially, the training objective isn't just "predicting the next word"; it is the structural likelihood of the observed bids. This forces the model to learn a low-dimensional representation ($z$) of the text that specifically captures features relevant to bidder valuations.
- Stage II (Structural Estimation & Inference): They freeze the language model and use the other half of the data to map the embeddings to the parameters of a Generalized Pareto Distribution (GPD). This distribution approximates the "tail" of the valuation distribution—the extreme upper end where the reserve price decision actually lives.
Key Innovation: The authors utilize a "count-free" identification strategy. By looking at the spacing between consecutive upper bids (e.g., the 2nd, 3rd, and 4th highest bids), they can identify the shape of the valuation tail without ever needing to know or model the total number of potential bidders.
Datasets
The study uses a massive transactional dataset from Bring a Trailer (BaT), a prominent US online auction platform for vehicles.
- Scope: Approximately 65,800 auctions between 2014 and 2023.
- Data Points: Unstructured listing descriptions, time-stamped and user-identifiable bids, and final sale outcomes.
- Filtering: The authors applied a "retention rule" to isolate auctions where bids credibly revealed true valuations (e.g., bids placed near the end of the auction).
Key Insights and Ideas
- Tail Information is Sufficient: You don't need to model the entire demand curve to set a reserve price. You only need to understand the very top of the valuation distribution. The GPD provides a mathematically robust way to approximate this tail from just a few top bids.
- Economic Alignment Matters: "Off-the-shelf" embeddings (pretrained for general semantic similarity) perform significantly worse than embeddings fine-tuned on the structural auction likelihood. Fine-tuning "directs" the model to ignore fluff and focus on value-driving descriptors.
- The "Mileage Gradient": The model successfully learned complex cross-sectional associations. For example, it correctly identified that the reserve price should decline as mileage increases, but it did so by looking at the text of the description rather than a structured "mileage" column.
- Invariance to Bidder Count: The optimal reserve price (under the Riley-Myerson framework) is invariant to the number of bidders. This allows the model to be operationally useful even when the "market size" for a specific niche item is unknown.
Key Results
- Revenue Gains: Counterfactual analysis suggests the model’s policy would have generated a net increase of $165.8 million, a 14.6% improvement over the incumbent reserve policy used by the platform.
- Conversion: Lowering reserves that were set too high would have converted $266.7 million in "failed" auctions into automatic sales.
- Accuracy: The fine-tuned model more than doubled the out-of-sample $R^2$ for the seller-value anchor (from 0.296 to 0.634) compared to a non-fine-tuned benchmark.
- Interpretability: Using "occlusion attributions" (deleting words to see how the reserve changes), the authors showed the model reacts logically: words like "911," "twin-turbocharged," and "2022" raised the recommended reserve, while "corrosion," "blemishes," and "dent" lowered it.
Limitations and Future Work
- Seller Valuation ($v_0$): The model requires the seller's own valuation as an input. In this paper, the authors calibrated this using a multiplier, but in practice, this remains a subjective policy input.
- Unitary Demand: The model assumes bidders are looking for exactly one item. While the authors tested for cross-auction overlap and found it limited, a more complex model might account for bidders who substitute between multiple concurrent auctions.
- Pretraining Data: As with all LLM research, there is a risk that the base model was exposed to some of the auction data during its original pretraining, though the authors' use of sample splitting for fine-tuning and inference mitigates this for the structural components.
- Bargaining Dynamics: The current model assumes a specific bargaining power structure if a reserve isn't met. Future work could explore how different post-auction negotiation rules might shift the optimal initial reserve.
Summary updated