The Fifth Elephant 2023 Winter

The Fifth Elephant 2023 Winter

On the engineering and business implications of AI & ML



Daood Shaheer


How many ads are too many? Enhancing ad serving to reduce campaign delivery variances

Submitted Sep 29, 2023


Ad serving refers to process of delivering online advertisements to users on a digital platform. It constitutes a crucial element in ad campaign management, as advertisers and platforms establish agreements regarding pricing and campaign behaviour. Consequently, optimizing under and over-delivery of ad impressions emerges as a critical concern for seamless and efficient delivery. Impression capping is a targeting option that limits how many times a user encounters a particular ad in each time frame. Conventional ad delivery systems only manage to achieve basic frequency capping (It is be defined as rendering only required number of impressions for the given duration of the campaign) wherein the impressions will always be lower than frequency cap but will not achieve the correct requirement of delivering the ‘X’ number of impressions. This leads to an inefficiency in the system wherein ad serving is unnecessarily trying to reach a wide audience of users whereas it did not need to do that.
Therefore, idea is to have models which accurately predict the number of impressions a user will view in each hour. Predicting user consumption patterns over time and directing campaigns towards user cohorts based on their consumption behaviours, it becomes possible to align with specific delivery objectives (Number of ad impressions, clicks, etc delivered over the entire campaign duration) which helps in reducing delivery variances.

Talk Outline

  • What is Impression capping, pacing, and what are the business impacts of the limitations of the current serving system?
  • What is forecasting, and how does a forecasting-driven model solve for the impression capping problem?
  • Solution Design
    • Why forecasting? Walkthrough of feature processing on user content consumption data
    • Analysing seasonality and trends in user consumption patterns
    • Navigating the aggregation problem: Signals aggregation and making predictions on user cohorts rather than of user-level prediction
    • What is Prophet, and what advantages does it offer over other forecasting models?
    • Walkthrough of how hourly and cohort-based predictions are utilised in the serving process
  • Offline and online evaluation strategies
  • Assessing the impact and summarizing the key takeaways


Daood Shaheer, Data Scientist (Glance)


{{ gettext('Login to leave a comment') }}

{{ gettext('Post a comment…') }}
{{ gettext('New comment') }}
{{ formTitle }}

{{ errorMsg }}

{{ gettext('No comments posted yet') }}

Hybrid access (members only)

Hosted by

All about data science and machine learning

Supported by


Providing all founders, at any stage, with free resources to build a successful startup.