Personalized Valuation Estimation for Bid Recommendation in Online English Auctions

Problem: Buyers in online English auctions must decide how much to bid without a reliable estimate of what the item is worth to them specifically. Existing platforms surface only the current auction price, offering no decision support. The core learning problem is estimating a buyer-specific, calibrated willingness-to-pay (WTP) distribution from behavioral marketplace data and using it to generate bid recommendations that update as the auction evolves. This setting differs from final-price prediction - which is item-level and non-personalized - and from auto-bidding for repeat-player advertisers, where the buyer has stable, measurable conversion objectives. Approach: We formulate the task as censored distributional regression. In proxy-bid English auctions, losing bidders reveal their maximum proxy bid directly (uncensored label), while winning bidders are right-censored at the final auction price, since their true maximum was never tested by competition. We train a gradient-boosted quantile regressor with a censored pinball loss at q ∈ {0.25, 0.50, 0.75, 0.90}. Three feature groups inform the model: (i) buyer history - category win rate, bid-to-price ratio, and a sniper index (fraction of bids placed in the auction's final decile); (ii) item attributes - category, condition, and comparable sold-price statistics from recently closed auctions; (iii) auction state - current price, unique bidder count, bid velocity, and time remaining. A low-rank buyer-category embedding, pre-trained on bid history and jointly fine-tuned with the regression head, enables graceful degradation to item-level prediction for cold-start buyers. Evaluation: We evaluate on historical completed auctions using a strict temporal hold-out split, with all buyer history features computed at auction start time to prevent leakage. On losing-bidder observations - where the ground-truth proxy bid is directly observed without simulation - we measure quantile coverage across q ∈ {0.25, 0.50, 0.75, 0.90}, expected calibration error, and median absolute deviation. Baselines include: the current auction price (floor), category-median sold price (non-personalized), buyer historical average proxy bid (personalized, no item context), buyer-agnostic quantile regression on item and state features only, and uncensored regression trained on winners only as a selection-bias control. Ablations isolate the marginal contributions of buyer personalization, auction-state features, and censoring correction. Context: Valuation estimation in English auctions is studied under the set-identification framework of Haile and Tamer [1], which derives WTP bounds under minimal rationality assumptions but does not address buyer-specific prediction or real-time auction dynamics. Structural econometric models [2] recover population valuation distributions under parametric assumptions without buyer personalization. ML work on auctions focuses on final-price prediction [3] or bid shading for repeat-player advertisers in first-price programmatic markets [4, 5] - where buyers have measurable conversion objectives rather than one-shot purchase decisions. Semi-parametric valuation estimation from second-price auction data [6] addresses aggregate distributions without personalization or auction-state conditioning. To our knowledge, no existing approach jointly models buyer-specific latent valuation, censoring structure, and evolving auction state for consumer-side bid recommendation in English auctions.