Budget Pacing Techniques and Their Application in 58.com Advertising Platform

Background – Cost‑per‑click (CPC) advertising relies on daily budgets; rapid budget consumption during traffic peaks leads to early ad shutdown and poor advertiser experience. Budget‑pacing distributes spend smoothly across the day, which is crucial for 58.com’s many small advertisers.

Main Budget‑Pacing Techniques – Two primary categories are discussed: (1) bid modification, where bids are adjusted based on current spend (citing Mehta et al., 2005); (2) probabilistic throttling, which allocates budget to time slots and adjusts exposure probability when consumption deviates from plan. The article also reviews reinforcement‑learning (RL) methods for real‑time bidding, referencing works by Cai et al. (2017) and Jin et al. (2018).

58.com Application Scenario – The platform serves four business lines (real‑estate, recruitment, yellow‑pages, used‑cars) with characteristics such as limited daily budgets, diverse product types, and complex traffic composition (PC, mobile, app, mini‑program). A pacing strategy based on traffic distribution was adopted, extending ad online time and lowering conversion cost.

Budget‑Pacing Based on Pacing Method – The approach consists of (1) budget allocation: dividing the daily budget into hourly slots according to traffic distribution; (2) budget control: using a PID controller to adjust the PTR (probability‑to‑request) value in real time. Initially a fixed‑rate adjustment was used; later it was replaced by a PID‑based method for better stability.

System Design – The pacing system comprises four modules: real‑time statistics (Flink + Kafka), budget allocation, PID control, and PTR filtering (including AB testing and blacklist handling). The workflow is illustrated in Figure 3.

Online Effects – A/B tests show that PID‑controlled pacing significantly reduces rapid budget depletion, increases ad online duration (e.g., from 13.8 h to 19.8 h), and lowers advertiser conversion cost. Figures 4 and 5 demonstrate the impact on ads with ample and extremely limited budgets.

Conclusion and Outlook – The pacing strategy has been successfully deployed across 58.com’s CPC products, achieving longer ad exposure and lower conversion cost. Future work includes combining pacing with other strategies (e.g., OCPC) and applying reinforcement‑learning for more intelligent budget management.

References – The article cites six papers covering generalized online matching, LinkedIn’s budget pacing, real‑time bid optimization, and reinforcement‑learning applications in display advertising.