[Paper Review] Optimal Real-Time Bidding Frameworks Discussion

Weinan Zhang, Kan Ren+1 more|arXiv (Cornell University)|Feb 2, 2016

Optimization and Search Problems1 references3 citations

TL;DR

This paper formulates a general optimal real-time bidding (RTB) framework that jointly optimizes for advertiser utility, budget constraints, and market dynamics. It derives closed-form bidding strategies for both first-price and second-price auctions using calculus of variations, showing that optimal bids are linear in utility under second-price rules and depend on market parameters under first-price rules.

ABSTRACT

This note is a complementary material for the solution of optimal real-time bidding function in paper "Optimal Real-Time Bidding for Display Advertising, KDD 2014", where the estimated cost is taken as the bid price, i.e., the upper bound of the true cost. Here we discuss a more general bid optimisation framework with various utility and cost functions.

Motivation & Objective

• To develop a unified mathematical framework for optimal real-time bidding in RTB that accounts for utility, budget constraints, and market price distributions.
• To generalize prior work that treated cost as an upper bound by modeling expected cost in second-price auctions more accurately.
• To derive optimal bidding functions using calculus of variations under both first-price and second-price auction mechanisms.
• To analyze how bid strategies are influenced by utility expectations, budget multipliers (λ), and market parameters (e.g., l in the price distribution).

Proposed method

• Formulates the bidding function b(r) as a mapping from estimated CTR r to bid price, with winning probability w(b) derived from market price distribution p_z(z).
• Uses the Lagrangian relaxation method to incorporate the budget constraint into the optimization, introducing a Lagrangian multiplier λ to balance utility and cost.
• Applies the Euler-Lagrange equation from calculus of variations to derive the first-order condition for optimal b(r), resulting in a differential equation involving u(r), c(b), w(b), and p_z(b).
• Solves the general optimality condition for two auction types: first-price (with cost c₁(b) = b) and second-price (with cost c₂(b) = E[z|z ≤ b]).
• For second-price auctions, derives a closed-form solution b(r) = u(r)/λ, showing linear dependence on utility and inverse dependence on λ.
• Solves for λ numerically by enforcing the budget constraint ∫c(b(r))w(b(r))p_r(r)dr = B/T, ensuring cost does not exceed budget.

Experimental results

Research questions

• RQ1 How can a general optimal bidding framework be constructed that jointly optimizes for utility, budget, and market price distribution in RTB?
• RQ2 What is the analytical form of the optimal bidding function under first-price and second-price auction mechanisms?
• RQ3 How do market parameters (e.g., l in p_z(z) = l/(l+z)²) and the Lagrangian multiplier λ influence the optimal bid price?
• RQ4 What is the impact of using expected cost (c₂(b)) versus bid price (c₁(b)) as a cost proxy in budget-constrained optimization?
• RQ5 How do utility, budget, and market information jointly shape the optimal bidding strategy in dynamic RTB environments?

Key findings

• Under second-price auctions, the optimal bid is b(r) = u(r)/λ, a linear function of the utility, with λ determined by the budget constraint.
• Under first-price auctions, the optimal bid depends on both the utility function and the market price distribution, with a closed-form solution b(r) = √(u(r)l/λ + l²) − l when p_z(z) = l/(l+z)².
• The Lagrangian multiplier λ controls the trade-off between utility maximization and budget adherence, and is solved numerically by equating expected cost to the budget per impression.
• Using the expected cost c₂(b) in second-price auctions leads to a higher risk of early budget exhaustion (up to 50%) compared to using the bid price as a conservative cost proxy.
• The market parameter l influences the winning probability w(b) and thus affects bid decisions, allowing adaptation to different market environments.
• The framework reveals that bid strategies are jointly shaped by advertiser utility, budget constraints (via λ), and market dynamics (via p_z(z)), with second-price bidding reflecting utility more directly than first-price bidding.