In the traditional landscape of digital advertising, the Second-Price Auction (SPA) reigned supreme. In an SPA, the winner pays the price of the second-highest bid, a mechanism that makes "truthful bidding"—submitting your actual valuation of the item—the mathematically optimal strategy.
However, as noted by Zhe Feng, Sebastien Lahaie, and colleagues in Reserve Price Optimization for First Price Auctions in Display Advertising, the industry has undergone a massive shift toward First-Price Auctions (FPA). In an FPA, the winner pays exactly what they bid. This change fundamentally breaks the incentive for truthfulness: if you bid your true value and win, your net utility is zero. To capture any surplus, you must strategically "shade" your bid—submitting a value lower than what the item is actually worth to you.
The Fundamental Trade-off
The core challenge of the FPA is a balancing act between two opposing forces:
- Profit Margin: Lowering your bid increases the surplus (value minus cost) you receive if you win.
- Win Probability: Lowering your bid decreases the likelihood that you will outbid your competitors.
As Yanjun Han, Zhengyuan Zhou, and others explain in Learning to Bid Optimally and Efficiently in Adversarial First-price Auctions, the "optimal" shade depends entirely on the distribution of competing bids. If you knew your opponent would bid \$5, you would bid \$5.01. Since you don't know their bid, you must estimate the "highest competing bid" distribution and pick a point that maximizes your expected utility.
Bidding Under Constraints
In real-world scenarios, advertisers don't just want to maximize surplus on a single auction; they operate under strict Budget and Return-on-Investment (ROI) constraints.
Gagan Aggarwal, Giannis Fikioris, and Mingfei Zhao address this in No-Regret Algorithms in non-Truthful Auctions with Budget and ROI Constraints. When a budget is limited, every dollar spent on a low-value auction is a dollar that cannot be spent on a high-value one later.
To solve this, researchers like Yige Wang in Adaptive Bidding Policies for First-Price Auctions with Budget Constraints under Non-stationarity use dual gradient descent. They treat the budget constraint as a "shadow price"—a dynamic multiplier that effectively tells the bidder: "Because our budget is running low, we should value our money more highly and shade our bids even more aggressively."
Learning and Stability
Because the optimal bid depends on what others are doing, an FPA is a moving target. If everyone shades their bids, the "price to win" drops, encouraging even deeper shading, until someone realizes they can win easily by shading less.
Xiaotie Deng and his team investigated whether these systems ever settle down in Nash Convergence of Mean-Based Learning Algorithms in First Price Auctions. They found that stability depends on the level of competition:
- If three or more bidders have the same high valuation, the auction reliably converges to a stable Nash Equilibrium.
- If only one bidder has a high valuation, the system can become permanently unstable, as that bidder constantly adjusts their shading to find the lowest possible winning price.
The Seller's Dilemma: Setting Reserves
The shift to FPAs also complicates life for the seller (the publisher). In an SPA, the seller can set a "reserve price" (a minimum floor) based on their own valuation. In an FPA, the seller must account for the fact that bidders are strategically shading.
In Reserve Price Optimization for First Price Auctions, Zhe Feng and Jon Schneider propose gradient-based frameworks to optimize these floors. The seller must predict how a change in the reserve price will cause bidders to re-calculate their shading. If the floor is too high, bidders might stop participating; if it's too low, the seller leaves money on the table that the bidders are "hiding" via shading.
Robustness and Machine Learning Hints
Modern approaches are moving toward making these strategies more robust to noise and uncertainty. Yanlin Qu and Aaron E. Flores introduced "Double Distributionally Robust" shading in Double Distributionally Robust Bid Shading for First Price Auctions, which protects bidders against errors in both their own valuation of the ad and their predictions of competitor behavior.
Furthermore, Aaron Flores and Wei Zhang have explored using "hints"—external predictions of the winning bid—in Leveraging the Hints: Adaptive Bidding in Repeated First-Price Auctions. They demonstrate that knowing the uncertainty of a competitor's bid prediction is often more valuable than the prediction itself, allowing bidders to shade aggressively when confidence is high and play it safe when the market is volatile.