AiBidder: A Cascaded Inference and Discretized Estimation Pipeline for Sparse E-Commerce Marketplaces

In e-commerce keyword auctions, automated bidding agents must dynamically balance relevance, conversion rate (CVR), and competitor pricing. While Reinforcement Learning is the standard formulation for determining bids, it is expensive to both train and serve at scale. We introduce AiBidder, a scalable system that circumvents this constraint by reducing bid selection over a discretized action space to a supervised estimation pipeline. However, standard regression on e-commerce conversions suffers from estimator collapse due to extreme zero-inflation. To solve this, AiBidder employs a Cascaded Inference Architecture: a dense proxy model first predicts Ad Impression, which then conditions secondary sparse estimators predicting Sales and Spend to yield Return on Ad Spend (RoAS = Sales / Spend). Evaluated on live production data, AiBidder substantially mitigates the estimator collapse seen in direct regression baselines. By prioritizing data efficiency and structural observability, the system successfully adapts to non-stationary markets.