Thursday, November 21, 2013

Break.com Case Study on ad serving.


Guillaume Roels of the Anderson UCLA created a 2008 case study on ad serving used in his MBA courses.

pdf_GR07.pdf

The case studies a 90 day period in the life of an adserver matching supply and demand.  There are three associated data elements:

  • Traffic data for the four inventory zones.  This gives expected "supply" on a daily basis in the form of ad unit requests (impressions) per zone per day on average.
  • A set of booked orders with flight dates, budget and impression allocation limits per zone.
  • A set of proposed orders with the same data elements as orders.
For convenience I have put the data from the case study in an excel file as well as CSV files.


There are many interesting questions to ask with the case:
  1. How much daily or weekly surplus/unsold inventory is available per zone?
  2. How many of the booked orders will be completed at a given level of supply?
  3. How much total revenue results from fully completed orders?
  4. How much total revenue results if partial credit is given for incomplete orders?
  5. How many of the proposals should be accepted such that they will complete?
  6. If the orders and proposals were treated equally which of the total set would complete?
  7. How much revenue results given either completed and/or partially complete orders?
  8. (advanced) Is there an optimal serving schedule (other than price order) that would result in more total revenue or more completed orders? 
While the MBA student might approach this with a spreadsheet analysis, writing code is likely a better approach. When implementing the simulation in code one has to make a few assumptions:
  • ASAP pacing of orders versus even pacing of volume per day across the flight dates.
  • Which traffic level to use per day (low/baseline/high).
  • The mechanism to decide the priority of orders to be served in a given day/zone.
This problem relates to both "matchmaking" and "mechanism design" in economics.  It is also a simple example of the problems one might solve in software in "computational advertising".  See below for a course with lecture slides touring the discipline.