狠狠撸

Про модели атрибуции
простыми словами

Your personal marketing analytics assistant
We help 20,000 analysts, marketing specialists and C-level executives to
manage data and make the right decisions on time.

Agenda
1. What is attribution?
2. What attribution models are available at the market?
3. Comparison of models
4. How to choose the model that will benefit your business?

What is attribution?
An example

A typical “customer 箩辞耻谤苍别测”...

Description
Why do we need attribution?
Interest / Awareness

Description
Consideration

Description
Conversion
Consideration

Description
Conversion
Retention
Consideration

Description
Conversion
Retention
Consideration
Less Targeted,
Less Attributable

Description
Conversion
Retention
Consideration
Less Targeted,
Less Attributable
Highly Targeted,
Highly Attributable

In the context of (online) marketing...

“Half the money I spend on
advertising is wasted; the
trouble is I don’t know which
half.”
- John Wanamaker, early 1900s

Questions business needs to answer
1. How do I achieve Sales plan?
2. How to allocate marketing budget?
3. How to decrease costs?
4. How to increase sales?

Position based models
First Click

First Click Last Click

First Click Last Click Last Non-Direct Click

Linear

Linear Time Decay

Linear Time Decay Position Based

Last Click attribution
Source: Ad Roll 2017
44% of marketers still use last click attribution.
Only 18% use algorithmic attribution.

72.4% of marketers indicate that they
● don’t know why they
chose their model
● selected the easiest
attribution option
available to them

Why is this happening???
1. Lack of understanding of the potential attribution impact
2. No ownership of attribution or analytics
3. Scattered data

Other Google Attribution models
Comprehensiveness
Actionability
Low
Low High
High
Attribution 360Google Attribution
Google
Analytics
Analytics 360
DCM
AdWords DS

What if not LNDC?
1. Markov Chains
2. Shepley value
3. Funnel Based model
4. Custom algorithms

If not Last Click & Last Non-Direct, than what?
1. You have data from different Ad platforms (needed point)
2. You want to estimate the value of every step and session for particular user
3. You want to understand which bundles of ad channels work well together

Methodology
1. How is the value distributed?
2. Where is it used?
3. What data you can (and have to) use?
4. Which question helps to answer?

Let’s start with figures and formulas?
Models:
1. Markov chains
2. Vector Shapley. The average contribution of all sources in the transaction
3. Funnel Based OWOX BI Attribution

Shapley value
Lets analyze tr1 = 500$ & tr2 = 300
facebook direct 500 USD
300 USDdirect
First transaction
Second transaction

Shapley value (for geeks)
Lets analyze tr1 = 500$ & tr2 = 300$
V1( {facebook} , {direct} ) = 500
V2( {direct} ) = 300
V3( {facebook} ) = 0
Ф1(facebook) = (1 - 1)! * (2 - 1)! / 2! * (0 - 0) + (2 - 1)! * (2 - 2)! / 2! * (500 -300) = 0 + 100 = 100
Ф2(direct) = (1 - 1)! * (2 - 1)! / 2! * (300 - 0) + (2 - 1)! * (2 - 2)! * (500 - 0)= 150 + 250 = 400

Example:
N = 2
● State Е
● State А
Probability matrix:
0.3 0.7
0.6 0.4
Markov chains

1 - Customer funnel 2 - How is it working 3 - Grouping by sources
C1 -> C2 -> C3 -> conversion (start) -> C1 -> C2 -> C3 -> (conversion) (start) -> C1, C1 -> C2, C2 -> C3, C3 -> (conversion)
C1 (start) -> C1 -> (null) (start) -> C1, C1 -> (null)
C2 -> C3 (start) -> C2 -> C3 -> (null) (start) -> C2, C2 -> C3, C3 -> (null)
Markov chains in Ecommerce
Let’s see 3 simple examples of clients’ behaviour:
C1 -> C2 -> C3 -> conversion
C1 -> unsuccessful conversion
C2 -> C3 -> unsuccessful conversion
С - session
(with particular source)

From To Probability General Probability
(start) C1 1/3 66.7%
(start) C1 1/3
(start) C2 1/3 33.3%
Total from (start) 3/3
C1 C2 1/2 50%
C1 (null) 1/2 50%
Total from C1 2/2
C2 C3 1/2 100%
C2 C3 1/2
Total from C2 2/2
C3 (conversion) 1/2 50%
C3 (null) 1/2 50%
Total from C3 2/2

For evaluation, we use the delete effect
P1 = (0,33 * 1 * 0,5) = 0,167
P2 = (0,33 * 0 * 0,5) = 0
P3 = (0,33 * 1 * 0) = 0
R1 = 1 - 0,167/0,33 = 0,5
R2 = 1 - 0 = 1
R3 = 1 - 0= 1
V1 = 0,5 / (0,5 + 1 + 1) = 0,2
V2 = 1 / (0,5 + 1 + 1) = 0,4
V3 = 1 / (0,5 + 1 + 1) = 0,4

Funnel Based OWOX BI
1. How is the value distributed?
2. Where is it used?
3. What data you can (and have to) use?
4. Which question helps to answer?

Step Users Probability Score Value
Visit 100.0%
Non-bounce visit 60.0% 60% 40 18%
Product page 42.0% 70% 30 13%
Add to cart 7.8% 19% 81 36%
Purchase 2.1% 27% 73 33%
224 100%
How is the value calculated?
Visit Non-bounce
Visit
Product
page
Add to cart Purchase
100%
60%
42%
7.8% 2.1%
60% 70% 19% 27%
18%
13%
36%
33%
= 40 / 224
= 30 / 224
= 8 / 224
= 73 / 224

But the funnel is not linear...

Comparison of how models work
Funnel Based Data-Driven (Analytics 360) Markov chains
1.Allows you to assess the mutual
influence of the channels on the
conversion and advancement along
the funnel
1.Allows you to estimate the mutual
influence of channels on the
conversion
1.Allows you to estimate the mutual
influence of channels on the
conversion
2.Allows you to find an inefficient
channel and tell where exactly it is not
effective. Resistant to nonlinearity.
2.Allows you to find an inefficient
channel. High accuracy of
calculations.
2.Evaluate which channel is the most
significant.
3.Underestimates the first step of the
funnel.
3. It does not evaluate progress on
the funnel, you can not connect offline
data from CRM
3.Underestimates the first link of the
chain, is unstable to the order in the
chains.
Answering the question:
How does the presence of a channel
affect conversion and when is this the
strongest influence?
How will the presence of the channel
affect the conversion?
How will the absence of a channel

Custom attribution models
Answear story:
1. Multi-brand online store selling clothes, footwear and accessories
2. Founded in Poland in 2010
3. Operates in several different counties

Divided channels in logical groups
● Comparison — price comparison services: hotline, ceneo.
● Affiliate — affiliate websites: zanox.
● Retargeting — retargeting services: criteo, rtbhouse.
● Cpc — paid search: google brand and non-brand + social.
● Display — display ads: google with graphic ads, viva.
● Email campaigns: external.

Defined main KPIs and assigned values

Assigned value to each channel

Attribution modeling 101, Mariia Bocheva

How did the presence of
the channel in the chain

How does the presence of a
channel affect conversion
and when is this the

How did the lack of a
channel in the chain

What indirect source
before conversion
was the last?
How did the lack of a
channel in the chain

Without a bidding integration,
attribution has no impact
Optimize marketing campaigns based on data
70% of marketers struggle to act
upon the insights of attribution.
Source: Ad Roll 2017

Key takeaways
1. Start with a clear strategy and set of objectives
2. Get internal buy-in for attribution
3. Focus on defining the customer journey
4. Consider physical as well as digital touchpoints
5. Ensure the data quality
6. Use flexible technology
7. Test different models that align with your business goals
8. Act on the results

Useful links
1. Comparison of different attribution models in the OWOX BI Blog
2. Custom attribution model by Answear
3. Article on how OWOX BI uses attribution for decision making
4. Markov Chains
5. Shapley Values

mail@owox.com
www.owox.com
Questions?

狠狠撸

Attribution modeling 101, Mariia Bocheva

More Related Content

Attribution modeling 101, Mariia Bocheva