top of page
Writer's pictureGabriel Gonçalves

Floor Price Optimization in Programmatic Advertising Using Machine Learning

Updated: Sep 25


Bid floor price optimization is like successfully clearing the bar in pole vaulting on the first attempt, by setting the right height.

Introduction

As the industry of programmatic advertising becomes more competitive more companies want to build their smart bidding software in-house Floor price optimization is an ML algorithm that offers publishers to maximize ad revenue by continuously adjusting the minimum acceptable bid for ad inventory. Using FPO publishers can better align with market demand and enhance their monetization efforts.


This article reviews the principles of floor price optimization, explores the approaches to implementing effective solutions, and provides an analysis of a case study from Yahoo and Forbes together with Amazon Ads. We also offer insights and feedback on how these strategies can be applied or improved upon, drawing from our expertise at TensorOps.


Understanding Floor Price Optimization

Floor prices, or reserve prices, are the minimum bids that a publisher or ad exchange is willing to accept for an advertising opportunity. Setting optimal floor prices can help SSPs maximize their revenues because it directly impacts both bid prices and inventory utilization.


Importance for Publishers

Publishers use bid floors as a way to communicate the advertisers the value of their inventory. When optimized bid floor can impact the publisher in to main ways:

  • Increased Bid Distribution: Optimal floor prices ensure that publishers are not undervaluing their ad space, thereby maximizing earnings from each impression.

  • Improved Fill Rates: By adjusting floor prices dynamically, publishers can better match supply with demand, leading to higher fill rates and reduced unsold inventory.


Risks of Over Optimization

However, over optimization can have negative impact on the publisher. Creating a very complicated algorithm can cause:

  • Market Confusion: Excessive or erratic changes in floor prices can confuse demand-side platforms (DSPs), potentially reducing bid participation and fill rates.

  • Predictability Issues: Overly complex algorithms may make pricing unpredictable, causing DSPs to divert their bidding strategies or shift budgets elsewhere.


Floor price issues

Strategic Approaches to Floor Price Optimization

Effective floor price optimization involves adjusting prices based on both market signals and internal data.


Market Signals

Here are a few factors that can lead to changes in the optimized bid floor:

  • Demand Fluctuations: Recognizing peak and off-peak hours to adjust prices accordingly.

  • Competition Levels: Monitoring competitor pricing and adjusting floors to remain competitive yet profitable.


Internal Data

When it comes to internal data publishers can rely on these parameters to estimate the optimal floor price.

  • User Quality and Ranking: Assessing user engagement metrics to determine the value of impressions.

  • Device and Demographics: Tailoring floor prices based on device types and user demographics that may yield higher engagement or conversion rates.

  • Psychographics: Leveraging behavioral data to understand user interests and purchasing intent.



Principles of Designing an ML-Based System for Floor Price Optimization

Implementing an effective floor price optimization system requires a divide and conquer approach that leverages machine learning to predict and respond to market dynamics. The design of such a system can be broken down into four key components:

  1. Continuous Data Collection through Randomized Floor Prices:

    "Playing" with the bid floor will impact the algorithms on the DSP, therefore you'd like to collect information about "what-if" cases. Allocating a small percentage of traffic to use randomized floor prices introduces controlled variability, which helps in mapping the current bid distributions and participation responses.

  2. Predicting Floor Price Impact on Participation:

    One component involves developing ML models that can accurately predict how changes in floor prices affect the participation rate of DSPs. By analyzing historical data, these models can estimate the likelihood of DSPs entering an auction at various floor price levels. Understanding this relationship is crucial because setting the floor price too high may deter DSPs from participating, leading to lower fill rates. Typically, simple models will be suitable for this phase, as the participation of the DSP is dependent more on the auction characteristics than on the bid floor itself. Therefore, accounting for the DSPs' previous participation in such auctions is ideal for a simple Bernoulli estimator with gradient boosted trees.

  3. Predicting Floor Price Impact on Bid Distribution:

    The second component focuses on forecasting how different floor prices influence the distribution of bids submitted by DSPs. ML models can capture patterns in bidding behavior, such as bid shading strategies and maximum willingness to pay. To model this you can assume a log-normal distribution and model it using neural networks.

  4. Optimizing for Profit Using Predictive Models:

    The final step involves integrating the predictive insights from the ML models into an optimization algorithm that seeks to maximize revenue. By simulating various floor price scenarios using the predicted participation rates and bid distributions, the optimizer can identify the floor price that is expected to yield the highest profit. This approach allows the system to "work against" the models, effectively finding the optimal balance between floor price levels, bidder participation, and bid amounts.


Yahoo's logic steps of the Floor Price Optimization (FPO)

Case Study Analysis: Yahoo's Floor Price Optimization Model

There are not many public cases about how companies implement their floor price optimization algorithms. This is why the article elaborating on Yahoo's transition to a first-price auction mechanism in 2019 reveals some interesting strategies. This transition allowed Yahoo to increase revenues while their model accounts for key industry characteristics and constraints, particularly those imposed by large bidders


Key Features of Yahoo's Model

  1. Compliance with Bidder Restrictions: The model respects bidder-imposed constraints such as:

    • Data Handling: Avoiding the use of intraday bid data to compute floors.

    • Uniform Floor Values: Fairing the game so that all bidders within the same type (regular DSPs vs. rebroadcasters) receive the same floor value.

  2. Auction Design Considerations: By focusing on first-price auctions, the model addresses the strategic bidding behaviors that differ from second-price auctions.

  3. Algorithmic Implementation: The model uses statistical tools and distribution functions (e.g., Weibull distribution) to estimate bidding behaviors and optimize floor prices.


Implementation and Results

  • Deployment: Initiated on display ad inventory in North America in June 2021, later expanding to other markets and video ad inventory.

  • Revenue Impact:

    • Display Inventory: Achieved an estimated annualized incremental revenue increase of +1.3%.

    • Video Inventory: Realized a revenue uplift of +2.5%.


Case Study: Forbes Media's Multi-Task Learning Model

Researchers at Forbes Media proposed a machine learning model that determines optimal reserve prices for individual ad impressions in real-time within first-price auctions. Their approach addresses several key challenges faced by publishers:


Challenges Addressed

The main challenges that the team had to face revolved both around modeling and data:

  1. Data Limitations: Publishers often lack access to detailed bid data, especially for underbid impressions where the highest bids are not visible.

  2. Risk of Exact Predictions: Predicting the exact highest bid is risky due to data noise and market volatility.

  3. Limited User Information: Publishers typically have less user data compared to advertisers, making it difficult to estimate user value accurately.


The solution

To address these challenges, the Forbes Media model utilizes a multi-task learning framework:

  • Main Task: Predicting the lower bound of the highest bid with a specified confidence level (e.g., 70% chance the actual bid will be above this lower bound).

  • Auxiliary Task: Predicting the failure rate of a given reserve price, i.e., the probability that the reserve price will not be met.


The main components of the solution are:

  • Interval Estimation over Point Estimation: Instead of predicting a single bid value, the model predicts a lower bound, acknowledging the uncertainty inherent in bid values.

  • Utilization of Both Outbid and Underbid Data: By leveraging both types of historical data, the model gains a more comprehensive understanding of bidder behavior.

  • Deep Neural Networks (DNNs): The model employs DNNs to capture complex interactions among features like user profiles, page content, ad placement, and contextual information.

  • Loss Functions:

    • Quality-Driven (QD) Loss Function: Used for predicting the lower bounds of bids, focusing on the width and coverage probability of the prediction intervals.

    • Cox Proportional Hazards Model: Applied for failure rate prediction, modeling the hazard (failure) rate of reserve prices.


Forbes Media didn't share the specifics of the impact on their revenues, however they did mention that it surpassed other method that were used before.


TensorOps Solution For SSPs

At TensorOps, we advocate for a data-driven approach to floor price optimization that leverages AI and machine learning. We also consider business implications of the applied algorithms.


Potential Risks and Mitigation Strategies

  • Over-Optimization: Continuous adjustments might lead to market instability. Implementing smoothing techniques or setting boundaries on the rate of change can mitigate this risk. With Google Ads for example, the optimization of the Floor Price can be treated as an "Experiment" and be managed explicitly.

  • Algorithm Transparency: Ensuring that the optimization process is transparent to clients can build trust and encourage continued participation from DSPs.


Steps for Effective Implementation

  1. Data Collection and Analysis: Build a big tolerance collector, Aggregate historical bid data, market signals, and internal performance metrics.

  2. Model Development: Utilize predictive models (e.g., reinforcement learning) to forecast bidder responses to floor price changes.

  3. Simulation and Testing: Run simulations to assess the impact of different floor pricing strategies on revenue and fill rates.

  4. Incremental Rollout: Implement changes gradually, monitoring key performance indicators (KPIs) to ensure positive outcomes.

  5. Continuous Learning: Incorporate feedback loops to refine models based on observed results.


Conclusion

Floor price optimization is a critical lever for publishers to maximize ad revenue in programmatic advertising. By intelligently adjusting floor prices based on market conditions and internal data, publishers can enhance revenue while maintaining healthy fill rates. Yahoo's case study provides valuable insights into practical implementation and the associated benefits.


 

At TensorOps, we are equipped to develop and implement advanced floor price optimization solutions tailored to the unique needs of publishers and ad-tech platforms. Our expertise in AI and machine learning positions us to drive innovation and deliver measurable results in this complex domain.


Comments


Sign up to get updates when we release another amazing article

Thanks for subscribing!

bottom of page