Effects of recency of interest group membership on the auction price

Hey,

In the spirit of adding data to decisions on thresholds of the specs and in
general adding more data to the conversation I wanted to highlight the
impact that we see in our network when it comes to what we call recency.

Recency can be defined easily as the duration in seconds since a user was
added to an interest group. In the current turtledove spec it says that the
ad bundle request for the interest group happens 4 hours after the browser
has been added to the interest group, this seems an arbitrary length of
time and I wanted to add some color on the consequence of such a long delay.

[image: Screen Shot 2020-06-29 at 3.12.15 PM.png]

This is a graph of normalized Bid CPM returned by our bidders vs how
recently the user was added to the list. As you can see, this value
degrades very very quickly. It's important to understand that this graph is
the result of a ML process aimed at optimizing for CTR and CPA, so this
curve is not the result of a manual attempt at gaming the metric, or an
a-priori notion of what traffic is valuable, this is learned.

To make this more clear, if one eliminates the aspect of the bid CPM, which
is a function of the auction as well, and simply focuses on how the ML
values the cookie vs recency equation:
[image: ValueVsRecency.png]
As you can see there is a significant drop in value that happens in the
first 6000 seconds since being added to the list which is very dramatic in
the first 1000 seconds.
I haven't matched this data yet against other vendors but I would be
extremely surprised if this wasn't true across the board. It's entirely
possible that a good chunk of the 50% drop in spend that Facebook and so
much other research has talked about is due in part to this effect.

Similarly, and this is a bit of an aside, similar pressure on prices is
caused by the various delays and limits throughout the specs currently
being evaluated, such as reporting delay.

Sadly a lot of the delays, noises and thresholds have very deterministic,
and significant, effects on the final prices advertisers would pay
publishers, so I was wondering if it wasn't time now to start to really put
some process in how these delays are established, or even if they are
necessary, since they carry a clear cost.

-- 
Valentino Volonghi
CTO, Founding Team
dialtone@nextroll.com