Source-linked AI summary
Real-time Bidding for Online Advertising: Measurement and Analysis
Shuai Yuan, Jun Wang, Xiaoxue Zhao
TL;DR
RTB introduced impression-level, user-centric bidding, but evidence about market infrastructure and bidding behavior remained limited. This paper empirically analyzes a production ad exchange using demand- and supply-side data, finding periodic activity patterns and substantial first-price payment under soft floors. The findings identify optimization issues involving temporal behavior, pacing, and ad-display frequency and recency.
Problem
The paper addresses limited empirical understanding of advertiser behavior and impression-level, user-centric bidding in RTB marketplaces.
Method
The authors analyze datasets from both demand and supply sides of a production ad exchange to measure infrastructure and bidding behavior.
Results
55.4% of total cost came from impressions bought in effective first-price auctions, despite the claimed second-price mechanism.
Takeaways & Limitations
Periodic patterns and soft-floor effects make time-dependent and more adaptive bidding optimization important, including pacing and frequency/recency considerations.
Takeaways & Limitations
The observed highest-bid distributions did not show strong log-normal properties, contrary to a widely adopted optimal-auction-design assumption.
Abstract
from arXiv · showhide
The real-time bidding (RTB), aka programmatic buying, has recently become the fastest growing area in online advertising. Instead of bulking buying and inventory-centric buying, RTB mimics stock exchanges and utilises computer algorithms to automatically buy and sell ads in real-time; It uses per impression context and targets the ads to specific people based on data about them, and hence dramatically increases the effectiveness of display advertising. In this paper, we provide an empirical analysis and measurement of a production ad exchange. Using the data sampled from both demand and supply side, we aim to provide first-hand insights into the emerging new impression selling infrastructure and its bidding behaviours, and help identifying research and design issues in such systems. From our study, we observed that periodic patterns occur in various statistics including impressions, clicks, bids, and conversion rates (both post-view and post-click), which suggest time-dependent models would be appropriate for capturing the repeated patterns in RTB. We also found that despite the claimed second price auction, the first price payment in fact is accounted for 55.4% of total cost due to the arrangement of the soft floor price. As such, we argue that the setting of soft floor price in the current RTB systems puts advertisers in a less favourable position. Furthermore, our analysis on the conversation rates shows that the current bidding strategy is far less optimal, indicating the significant needs for optimisation algorithms incorporating the facts such as the temporal behaviours, the frequency and recency of the ad displays, which have not been well considered in the past.
1. INTRODUCTION
RTB reshapes display advertising around impression-level, user-centric auctions mediated by ad exchanges, DSPs, and SSPs. The paper introduces this market structure and studies its empirical and research challenges using data from both demand and supply sides.
- Paper scope: The paper empirically analyzes a production ad exchange using datasets from both demand and supply sides to examine bidder behavior and RTB research challenges.The study focuses on impression-level and user-centric bidding rather than bulk buying and inventory-centric buying.
- Market evolution: Ad exchanges emerged to address supply-demand imbalance and unsold impressions in ad networks.Advertisers previously managed multiple ad-network channels to find affordable or sufficient inventory.
- RTB market structure: RTB queries DSPs for bids on each impression, optionally using third-party user data before returning bids through the exchange.The winning ad is then displayed to the targeted user on the publisher’s website.
- Market structure: DSPs centralize advertiser participation and enable finer-grained, higher-frequency impression optimization, while SSPs provide publisher-side yield tools such as reserve prices and bidder preferences.These platforms reduce the need to manage multiple ad-network registrations and support publisher revenue optimization.
- Open research issues: The current marketplace lacks widely adopted links between SSPs and data exchanges and among ad exchanges, limiting interconnection opportunities identified by the authors.The authors suggest that publisher access to user data and exchange-to-exchange impression sharing could support revenue optimization and a unified marketplace.
2. THE EMPIRICAL STUDY OF RTB
The empirical study reports general bidder and bid statistics, then examines bidding behavior alongside frequency, recency, and associated optimization problems. Its analysis centers on floor-price detection, daily pacing, and selective bidding.
- Study organization: The study first reports dataset statistics and general analyses of bidders and bids.
- Study organization: It then examines bidding behavior and the effects of ad-display frequency and recency.These factors are linked to problems including floor-price detection, daily pacing, and selective bidding.
- Research problems: The section frames floor-price detection, daily pacing, and selective bidding as associated problems requiring analysis.
2.1 Dataset and the System’s Properties
The production ad-exchange study uses demand- and supply-side samples to examine periodic activity and auction payments. It finds daily conversion patterns and substantial effective first-price payment associated with soft floor prices.
- Dataset: 52,850,635 impressions, 72,958 clicks, and 37,978 conversions were sampled from advertiser logs, while 12,965,119 auctions covered 50 publisher placements.The demand-side sample spans February–May 2013; auction logs span 14 December 2012–24 May 2013 across 16 websites.
- Periodic Patterns: Daily patterns appear in impressions, clicks, conversions, and conversion rates, with daytime showing more conversions and higher CVRs than sleeping hours.Post-view and post-click conversions were plotted separately.
- The Impact of the Soft Floor Price: A high soft floor price can convert the auction from second-price to first-price payment, charging the winner its bid; advertisers may not know the soft floor exists.This differs from the hard floor, which is the traditional reserve price.
- The Impact of the Soft Floor Price: 40% of impressions were bought through effective first-price auctions, accounting for 55.4% of total cost.The study classified cases where bid price equaled price paid as first-price auctions.
- The Impact of the Soft Floor Price: The authors argue that soft-floor settings put advertisers in an unfavourable position and warrant exploration of publisher-specific floor prices.In high-soft-floor scenarios, a winner may lower its bid while still winning to reduce cost.
2.2 Bidding Behaviours
The study examines time-varying bidding and daily budget pacing in RTB. It finds weak support for log-normal bid assumptions, supply–demand timing imbalances, and shortcomings in widely used pacing strategies.
- Bids’ Distribution: Winning-bid averages peak around 6–8am, when impressions are lower and bidder counts are higher.Bid variation throughout the day motivated splitting auctions by hour-of-day.
- Bids’ Distribution: Less than 1% of fitted bid distributions passed either the Shapiro-Wilk or Anderson-Darling test, regardless of hourly bidder or impression counts.The fits considered highest bids, highest plus second-highest bids, and all bids for placement-hour tuples.
- The Daily Pacing: No pacing can deplete budgets prematurely, while uniform pacing can miss early high-quality impressions or under-deliver when late traffic is insufficient.Both strategies are described as widely used but inadequate for varying traffic quality and volume.
- The Daily Pacing: Bidder counts peak before impressions, indicating hourly supply–demand imbalance; early competition raises winning bids without necessarily improving performance because CVRs peak in the evening.The authors therefore characterize intense early bidding as unreasonable.
- The Daily Pacing: Dynamic pacing against performance is proposed as an alternative that spends more during hours generating more clicks or conversions.The allocation problem also requires revising remaining budget after each time step because of practical latency.
2.3 Conversion Rates and Selective Bidding
Conversion efficiency depends on campaign-specific frequency and recency settings, so selective bidding should account for both exposure counts and elapsed time. The paper shows that mismatched caps can forfeit conversions or waste impressions and that RTB enables user-level control.
- The Frequency Factor: Different campaigns require different frequency caps: campaign 1 peaks at 6-10 impressions, whereas campaign 2 peaks at 2-5.For campaign 1, a 2-5 cap would miss most conversions; for campaign 2, a 6-10 cap could waste nearly half of impressions.
- The Frequency Factor: Frequency-cap evaluation depends on the objective: CVR and CPA favor FC=3, total conversions favor FC=5, and profit-and-cost favors FC=2.The comparison assumes 100 users, CPM=10, and a conversion goal worth 500.
- The Frequency Factor: FC=2 appears most profitable, while FC=3 may provide more conversions at low cost and support future conversions.The paper presents FC=3 as reasonable when considering longer-term impact.
- The Recency Factor: Recency settings must reflect campaign behavior because intensive advertising can be inefficient for some campaigns, while rapid conversion campaigns require more intense advertising.The paper gives financial services and flight booking as contrasting examples.
- The Recency Factor: Both campaigns reach their highest CVR at 1-5 minutes, but campaign 2 retains non-negligible CVR after 14-30 days.A strict long-term recency cap may lose conversions for campaign 2, whereas a loose cap may waste campaign 1’s budget.
- Selective Bidding: RTB can apply individual frequency and recency caps to decide whether to bid on a specific impression.The paper identifies these settings as important for advertising efficiency and notes that they can be set at finer granularity than the creative level.
3. CONCLUSION
The paper uses production ad-exchange data to examine demand-side bidding and related RTB research challenges. It identifies unresolved problems in floor-price detection, daily pacing, and frequency/recency settings, while leaving algorithm development and evaluation for future work.
- CONCLUSION: The study analyzes advertiser and delegate behavior in a production ad exchange, focusing on impression-level and user-centric bidding.
- CONCLUSION: Floor-price detection, daily pacing, and frequency/recency setting remain unaddressed problems identified through dataset analysis.
- CONCLUSION: Algorithm development and evaluation for these problems are left to future work.