Today, a large part of advertising is conducted online. This automated and optimized process typically occurs through real-time bidding auctions. Accurate win rate estimation is crucial to optimize the bids with which we participate in these auctions. Although traditional pointwise models for win probability achieve high accuracy, their predictions are mutually independent, meaning that they may not increase monotonically with respect to the bid amount. In this work, we explore a transition to parametric curve prediction using a sigmoid function. We developed and evaluated five consecutive model iterations with the aim of systematically eliminating gaps in prediction accuracy.
Results show that the parametric approach lags behind the baseline model in terms of relative information gain. However, through feature optimization and modifications of function implementation, we successfully reduced the initial 5% difference to just 2%. The main finding of this study is that improvements in standard point metrics in theory do not necessarily reflect the actual quality of the predicted curve in practice. Consequently, these results offer an opportunity for a direct evaluation of the model in a production environment using optimization procedures (such as first-price optimization), which will demonstrate its actual business impact.
|
