Market Rate Estimation

What should I bid in order to win in the RTB auction? Traders often spend hours every day trying to figure out the answer to this question.

Typically the process goes like this:

Step 1) When initially setting up a campaign, traders guess how much it will cost to win impressions and meet their delivery goals. They usually start fairly low base bid - say $1 CPM.
Step 2) They run their campaign for a day and then look at analytics to see how much inventory they’ve won.
Step 3) If their analytics tell them they aren’t winning enough impressions to hit their delivery goal and their win rate is low, they increase their bid – say to $1.50, and then repeat steps 1-3, guessing until they bid enough to win the auction.

This “guess and check” process for estimating how much it will cost to win in the RTB auction has two pretty obvious drawbacks:

1) If the trader guesses too low – they may not win any impressions at all, and never get the chance to learn on a specific piece of inventory
2) If the trader guesses too high – they may be subject to publisher soft floors and end up overpaying for inventory they could have otherwise won for less money.

However, there is also a third, less obvious drawback as well:

3) The price required to win an RTB auction on any particular slide of inventory is not static throughout the entire day. It can actually fluctuate up to 400%.

The market price to win a specific piece of inventory usually looks like this:

alt

Supply of digital ad impressions fluctuates wildly throughout the day. At night when more people are sleeping, fewer people are surfing the web and visiting websites so there is less supply of online ad impressions. Somewhat counter intuitively, because there aren’t as many impressions available, online ad inventory is more expensive to buy at night (low supply, high price). During the day, when people are awake, there are many more impressions available, and those impression are generally less expensive. Bidding a fixed price throughout the entire day actually underpays for inventory when supply is low (at night) and overpays for inventory during the day (when supply is high).

Here is an actual snapshot of the comparison Average CPM and Average Impression volume by hour of day in Japan.

As you can see – it’s more expensive to buy impressions at 5 o’clock in the morning than it is at 12 noon.

alt

Market Rate Estimation is a calculation of what price bid it will take to win any given piece of inventory at any given time across the entire landscape of online ad impressions.

Here’s how it works:

We break down the entire landscape of online ad impressions into buckets comprised of unique combinations of Tag;Hour;Geo;Size;Day/HourofWeek. For each of those buckets we predict in real time what price is going to win different percentages of inventory (e.g. 10%, 50% or 80%). The prediction works by looking back at historical prices of what it took to win that inventory in the past and then adjusts up or down based on the data we’ve already collected from the current day.

Current tests indicate that our prediction is 96%-97% accurate.

How can traders turn this theory into practice? Console offers the option to “Target Reach and Delivery” and select either EAP or ECP as a bidding strategy.

You can also use market rate estimation as a variable inside APB. Simply insert “estimated_average_price” into your APB tree to dynamically bid a price that will win 50% of auctions or insert “estimated_clearing_price” into your APB tree to dynamically bid a price that will win 80% of auctions. For a deeper dive of how this works, see the Developer Blog article "Using MRE in Bonsai Smart Leaves

Andrew Eifler leads the Product Line Management team for our Advertiser Technology Group. Written in conjunction with Adam Petranovich and Abraham Greenstein from the AppNexus Data Science team.

Andrew Eifler

Read more posts by this author.

Subscribe to AppNexus Product Blog

Get the latest posts delivered right to your inbox.

or subscribe via RSS with Feedly!