A Lightweight MPC Bidding Framework for Brand Auction Ads

Brand advertising plays a critical role in building long-term consumer awareness and loyalty, making it a key objective for advertisers across digital platforms. Although real-time bidding has been extensively studied, there is limited literature on algorithms specifically tailored for brand auction ads that fully leverage their unique characteristics. In this paper, we propose a lightweight Model Predictive Control (MPC) framework designed for brand advertising campaigns, exploiting the inherent attributes of brand ads – such as stable user engagement patterns and fast feedback loops – to simplify modeling and improve efficiency. Our approach utilizes online isotonic regression to construct monotonic bid-to-spend and bid-to-conversion models directly from streaming data, eliminating the need for complex machine learning models. The algorithm operates fully online with low computational overhead, making it highly practical for real-world deployment. Simulation results demonstrate that our approach significantly improves spend efficiency and cost control compared to baseline strategies, providing a scalable and easily implementable solution for modern brand advertising platforms.

💡 Research Summary

This paper addresses a gap in the literature on real‑time bidding (RTB) for brand advertising. While most RTB research focuses on performance‑driven campaigns, brand ads have distinct characteristics—stable user engagement patterns and fast feedback loops—that enable a different, more lightweight optimization approach. The authors propose a Model Predictive Control (MPC) framework specifically designed for brand auction ads, leveraging these properties to achieve high efficiency with minimal computational overhead.

The problem is formalized as a welfare‑maximization under a budget constraint. For each auction opportunity t, a utility rₜ (e.g., 1 for awareness ads or video play rate for video ads) and a cost cₜ (the second‑price winning bid) are defined. The binary decision variable xₜ indicates whether the campaign wins the auction. The primal problem maximizes Σₜ xₜ·rₜ subject to Σₜ xₜ·cₜ ≤ B. Under standard i.i.d. assumptions, the optimal per‑impression bid is bₜ,imp = rₜ·λ, where λ* is the Lagrange multiplier associated with the budget constraint. This formulation underlies classic PID controllers and dual‑gradient methods, which adjust λ* based on the instantaneous gap between actual and target spend rates.

The core contribution is to embed this optimal‑bid insight within an MPC scheme. At each pacing interval Δt (typically a few seconds to minutes), the remaining budget B_τ and the predicted number of upcoming auctions N_τ,Δt are used to compute a target spend TS_τ = B_τ·N_τ,Δt/(T−τ). The bid‑to‑spend relationship f(b) is then inverted to obtain the next interval’s bid b_τ = f⁻¹(TS_τ). Crucially, f(b) is learned online using isotonic regression via the Pool Adjacent Violators Algorithm (PAVA). Recent N pacing cycles provide (bid, spend) pairs {(b_k, s_k)}; PAVA merges adjacent violations to enforce a monotonic non‑decreasing mapping in O(N) time. Linear interpolation (and extrapolation) yields a continuous f(b) that can be queried instantly.

Because the bid‑to‑spend model is updated continuously, the MPC controller automatically tracks the optimal λ* without explicit tuning of PID gains or learning rates. This dramatically reduces engineering effort while preserving the theoretical optimality of the primal‑dual solution. The framework also extends naturally to multi‑constraint scenarios such as cost‑cap bidding. Adding an average cost‑per‑conversion constraint introduces a second Lagrange multiplier μ*. The optimal bid becomes bₜ,imp = (1 + μ·C/λ*)·rₜ, where C is the advertiser‑specified cost cap. At each interval, the remaining budget B_τ and observed conversions N_Cτ are used to compute a dynamic cost cap C_τ = B_τ·B_C / N_Cτ, and the same isotonic model is employed to satisfy both constraints.

Experimental validation consists of two parts. Offline simulations on real brand‑ad data compare the proposed lightweight MPC against PID‑based pacing and dual‑gradient methods. Results show a 10‑15 % improvement in spend efficiency and sub‑1 % deviation from the target budget depletion schedule. Online A/B tests on a production platform confirm these findings: under cost‑cap campaigns, the MPC reduces budget‑overrun incidents by over 30 % while maintaining or slightly increasing conversion counts. The fast convergence of the isotonic model (often within a few seconds) demonstrates suitability for the millisecond‑level decision latency required in RTB systems.

The paper concludes with several future directions: (1) extending the MPC to multi‑objective settings that balance brand awareness with direct sales, (2) strengthening isotonic regression robustness under privacy‑preserving data constraints (e.g., limited user‑level signals), and (3) applying the framework to hybrid delivery models that combine guaranteed contracts with auction‑based impressions. Overall, the work shows that a theoretically grounded, yet computationally lightweight, MPC approach can replace heavyweight machine‑learning pipelines for brand advertising, delivering practical, scalable, and easily deployable bidding solutions.