Billions and Billions: Machines, Algorithms, and Growing Business in Programamtic Markets

Billions and Billions:
Machines, Algorithms, and Growing
Business in Programmatic Markets
Ari Buchalter COO, MediaMath
PhD, Astrophysics

What do these things have in common?
The Digital Advertising Universe The Actual Universe
• Both are complex systems
• Math can be applied to understand both

The evolution of media decision making
Past 60 years – “Audience-based” Today – “Goal-based”
Describe your audience
Figure out media they consume
Buy placements, wait, and hope
Get report, manually adjust
Define your marketing goal
Capture all the data (media, user)
Model it to identify what works
Automate the buying
Humans making coarse
decisions based on proxies,
averages, and indexes
Machines making exact
decisions based on granular
user data

We’ve seen this movie before…

…and we understand the benefits
Buy in batch (wheat + chaff) Buy what you want (wheat only)
Fixed price, regardless of value Variable bidding, aligned with value
Little/no insight into true drivers Full insights into “what” & “why”
Manual, labor-intensive (~5/FTE) Fully automated, scalable (~50/FTE)
And the results are typically 10x better, BUT there’s a cost….
Analyze 10-20 buys weekly Analyze 1MM opps. per sec.

Let’s talk about Big Data in Programmatic
~100 BILLION impressions per day
~100 variables per impression
~100 values per variable
EQUALS
~1,000,000,000,000,000
Possible combinations of data per day
(1015 = ONE QUADRILLION)

Making sense of the chaos
Algorithms
Optimization
Programmatic
Automation
Predictive Modeling
Machine learning
Decision engines

Getting inside the RTB transaction
SSP or
Exchange
Publisher
Consumer
DSP
Advertiser
Agency or
Trading Desk

The two (buyer) questions that matter
What is the right bid for each impression?
Which impressions should I buy?

Why does question #1 matter?
Too high
Overpay &
underperform
Too low
Lose out &
underspend
“Goldilocks” bid
Maximize scale &
performance

Why does question #2 matter?
• ~$100MM/day of RTB supply
• Typical campaign spends ~$1K/day
(0.001% of total supply)
• Not buying the RIGHT 0.001% is
throwing money away

Answering the questions ain’t easy
Data is large, and growing  need technology at scale
It’s called different things  need to “normalize” data
Data interactions are complex  need sophisticated models
Mix of goals (upper/lower funnel)  need flexible methodology
Supply & demand constantly changing  need to remodel often
Clients need to understand  need intuitive, transparent output
It’s all in real-time (100ms)  need speed without latency
Only a machine-learning algorithmic approach can handle
the size, variability, complexity, and speed required

Question #1 – A simple exercise
$1 prize
Flip a coin to
win 1 dollar
50% chance $0.50
Bid Price
Goal Value x Action Rate = Bid Price

Question #1 – The real thing
Goal Value x Action Rate = Bid Price
1% chance consumer
takes desired action
(purchase)
$50 value to
advertiser
(CPA)
Bid for an
RTB ad
$0.50 bid
price
(breakeven)
YOUR AD HERE

Question #1 – The Goal Value
(Input) (Prediction) (Output)
Goal Value
x Action Rate = Bid Price
The goal can be anything at all:
• Branding: positive survey response (awareness, intent, etc.)
• Engagement: site visit, site action (locate store, post comment)
• Conversion: signup, application, purchase, etc.
• Retention: repeat purchase, renewal, upsell
If it can be measured, it can be made better by math

Question #1 – The Action Rate
(Input) (Prediction) (Output)
Goal Value x Action Rate
= Bid Price
Predictive modeling: the process by which a mathematical model is created to
predict the probability of an outcome, usually based on historical input data
The model should base the prediction on all available data:
• User: site activity (1p), interests & behaviors (3p), geo, TOD, DOW, etc.
• Media: channel, publisher, page, ad size, above/below fold, etc.
• Creative: image, offer, call to action, etc.

Answering question #1
Video
Publisher: YouTube
Unit: 15 sec pre roll
Time: 16.46 – 17.00
Age: 25-34
Gender: Male
Price: $15.76 CPM
Social
Publisher: FBX
Unit: Newsfeed
Day: Tuesday
Time: 5.00pm – 5.15pm
Price: $2.30 CPM
Display
Publisher: Rubicon
Data: Rakuten Male
Location: Tokyo
Creative size: 160 x 600
Price: $0.63 CPM
A different model for every creative in every
campaign of every advertiser – all in real time!

Question #2 – Which ones to bid on?
Optimization: the process of making
the best choice among a set of
options to achieve a desired goal,
usually under some constraints
Example – Shopping for food
Constraints: fixed budget,
nothing artificial
Goal: Most mass of food?
Most volume of food?
Healthiest mix?

Question #2 – Two important concepts
1) Bid Price: How much the impression is worth to the buyer
• Depends on who the publisher is and who the advertiser is
• Is a measure of quality (i.e., what it’s worth to the buyer)
2) Market Price: The price the impression will clear for
• Depends on the entire marketplace
• Also obtained through predictive modeling

Question #2 – A meaty example
Bid Price:
$30
Bid Price:
$30
$30
High
(good quality)
Bid Price
Low
(poor quality)
Bid Price:
$2
Bid Price:
$2
$2
YES!
Market Price
High
Eh, OK
(not a deal)
Low
(a deal!)
NO!
Eh, OK
$30
Selling for:
$30
Selling for:
$30
$2
Selling for:
$2
Selling for:
$2

Answering question #2 – Which to bid on?
III IV
Quality-driven
performance
<10% of impressions
I II
performance
40-70% of impressions
Value-driven
performance
<5% of impressions
Cost-driven
performance
20-50% of impressions
Relative Value
Low High
Non
Low High
Bid Price

Putting it all together
1) Use a predictive model
to determine what each
impression is worth
2) Use optimization to
determine which
impressions to bid for

So where do I get me some of those?
Find a partner who:
 Leverages robust technology – ask to see the scale & speed
 Has proven results – across verticals, geographies, over time
 Will expose the “black box” – transparency & insights are key!
 Has cross-channel capabilities – display, video, social, mobile, premium, BYO
 Has broad integrations – 3p data, surveys, viewability, attribution, etc.
 Can incorporate 1st party offline data – increasingly important
 Develops custom solutions – to suit your unique business needs
 Makes it easy – execution, workflow, reporting, testing, etc.
 Provides thought partnership & great service – it’s not just machines!
(machines just enable people to do the REAL value-added work)

Forrester DSP Wave: MediaMath is #1
“MediaMath boasts excellent algorithmic
optimization capabilities (including a
multifaceted view of the decisioning engine’s
output), and its multichannel media and data
access is both broad and deep.”
“MediaMath is a great all-around choice
for buyers in market for a DSP.”
“Its large employee base and diverse, well-tenured
management team also provide the
necessary foundation for it to execute
effectively on its strategic vision: to
empower marketing professionals with a
flexible, easy-to-use, multichannel platform.”

Billions and Billions: Machines, Algorithms, and Growing Business in Programamtic Markets

More Related Content

What's hot

Viewers also liked

Similar to Billions and Billions: Machines, Algorithms, and Growing Business in Programamtic Markets

More from MediaMath

Recently uploaded

Billions and Billions: Machines, Algorithms, and Growing Business in Programamtic Markets