40

Lucie Sperkova - Pioneering multi-channel attribution for the lack of comprehensive solutions

Embed Size (px)

Citation preview

Why do we attribute

(1838 - 1922)

Why do we attribute

Budget Allocation

Media Plan

Channel Performance and Value

Customer Journeys

Data Source: Google Analysis <Marketing Attribution: Valuing the Customer Journey>

Separate silos

SEARCH AD BUDGET

SEARCH CLICKS &

IMPRESSIONS

SEARCH CONVERSIONS

DISPLAY AD BUDGET

DISPLAY CLICKS & IMPRESSIONS

DISPLAY CONVERSIONS

PROGRAMMATICPPC

SOCIAL AD BUDGET

SOCIAL

SOCIAL CLICKS &

IMPRESSIONS

SOCIAL CONVERSIONS

Uniform data

ADVERTISING BUDGET

CLICKS & IMPRESSIONS

CONVERSIONS

PROGRAMMATICPPC SOCIAL

Why own solution?

all impressions

full browsing history

paths which did not make conversion

cross-device

paths in their whole length (Google cuts them to 4 channels)

sophisticated methods

CRM data

Last-click (heuristic) problem

more information in: John Murphy, 2014

Last-click (heuristic) problem

Last-click (heuristic) problem

First-click (heuristic) problem

Display campaigns matter

Channel dynamics

Customer behaviour (consumer funnel) matter

Was the conversion caused by this channel?

logistic regression models (Shao & Li 2011; Klapdor 2013)

Data-driven models

logistic regression models (Shao & Li 2011; Klapdor 2013)

game theory-based models (Berman, 2015; Dalessandroet al. 2012)

Bayesian models (Li & Kannan 2014; Nottorf 2014)

mutually exciting point process models (Xu, Duan, & Whinston, 2014)

hidden Markov models (Abhishek, Fader, & Hosanagar 2015; Anderl et al. 2014)

Data-driven models

logistic regression models (Shao & Li 2011; Klapdor 2013)

game theory-based models (Berman, 2015; Dalessandroet al. 2012)

Bayesian models (Li & Kannan 2014; Nottorf 2014)

mutually exciting point process models (Xu, Duan, & Whinston, 2014)

hidden Markov models (Abhishek, Fader, & Hosanagar 2015; Anderl et al. 2014)

VAR models (Kireyev, Pauwels, & Gupta 2016)

multivariate time-serie models (Anderl et al. 2015)

survival models

Data-driven models

Simple Probabilistic Method Shao and Li, 2011

Shapley Value Aspa Lekka, 2014

Hidden Markov Model Anderl et al., 2014

Science behind the models

Criteria / ModelHeuristic Simple probabilistic Shapley value Markov

Objectivity and fairness No Yes Yes Yes

Predictive accuracy No Partly - Yes

Carryover and spillover effects No Partly Yes Yes

Data-driven No Yes Yes Yes

Interpretability Yes Yes Partly Partly

Customers’ heterogeneity No Partly Partly Yes

Robustness No Partly - Yes

Algorithm efficiency YesSatisfactory for lower

ordersNo

Satisfactory for lower orders

Versatility Yes Yes Yes Yes

Criteria / ModelHeuristic Simple probabilistic Shapley value Markov

Objectivity and fairness No Yes Yes Yes

Predictive accuracy No Partly Yes Yes

Carryover and spillover effects No Partly Yes Yes

Data-driven No Yes Yes Yes

Interpretability Yes Yes Partly Partly

Customers’ heterogeneity No Partly Partly Yes

Robustness No Partly Yes Yes

Algorithm efficiency YesSatisfactory for lower

ordersNo

Satisfactory for lower orders

Versatility Yes Yes Yes Yes

“We have no place to grow; PPC campaigns has used up its potential.”

“Effective revenue share is smaller than was the goal so that we could spend more money, but it was not where to spend… We put more money to Google in

Slovakia market, and ERS got even cheaper.”

How to get from last-click trap

Methodology

Our clients are heterogeneous, but we have to be able to maintain uniform solution.

Data collection

Data pre–processing Run models Budget

reallocationResults testing and validation

Descriptive analysis

Data cleaning

Data selection

Pathsreconstruction

Technology:

Data Collection

Data collection all raw data including all clicks, impressions, web entrances

Data granularity channel - campaign - media - placement

Channels free channels are taken into account

Data preparation: 80% success

Data cleaning exclude robotic transactions exclude disabled cookiesexclude not visible impressionsexclude repeated actualisations of websitescombine impressions in 30-minute interval

Transformation to journeys

non-conversion taken in accountexclude paths longer than treshold

Data: > 1,5 TB Rows: > 3,2 billions

Consecutive impressions

visualisation: Crossmasters

Consecutive impressions

visualisation: Crossmasters

Consecutive impressions

visualisation: Crossmasters

Modelling

Period analysed monthly basis

Time window 1 month

Reporting

CPA

ROAS (%)

channel cost

number of channel

conversions

channel weight

channel cost weight

ROAS > 100 % channel is undervalued

channel cost weight = channel cost

sum of all cost

Proposed Budget

actual budget * ROAS=

=

=

Return of advertising spends (ROAS) channel weight

channel cost weight

channel cost first-click last-click linear-touch shapley value

simple probabilistic

markov

SKLIK 193 000 CZK 162 % 194 % 165 % 147 % 193 % 35 %

RTB 13 000 CZK 110 % 70 % 90 % 70 % 60 % 189 %

RTB 2 10 000 CZK 1057 % 319 % 663 % 561 % 369 % 14 %

Proposed Budget = actual budget * ROAS

channel cost first-click last-click linear-touch shapley value

simple probabilistic

markov

SKLIK 193 000 CZK 313 000 CZK 375 000 CZK 320 000 CZK 284 000 CZK 373 000 CZK 68 000 CZK

RTB 13 000 CZK 15 000 CZK 9 000 CZK 12 000 CZK 10 000 CZK 8 000 CZK 25 000 CZK

RTB 2 10 000 CZK 108 000 CZK 33 000 CZK 69 000 CZK 57 000 CZK 38 000 CZK 148 000 CZK

... ... ... ... ... ... ... ...

... ... ... ... ... ... ... ...

... ... ... ... ... ... ... ...

SUM 1 300 000 CZK 1 300 000 CZK 1 300 000 CZK 1 300 000 CZK 1 300 000 CZK 1 300 000 CZK 1 300 000 CZK

Proposed Budget = actual budget * ROAS

channel cost first-click last-click linear-touch shapley value

simple probabilistic

markov

SKLIK 193 000 CZK 313 000 CZK 375 000 CZK 320 000 CZK 284 000 CZK 373 000 CZK 68 000 CZK

RTB 13 000 CZK 15 000 CZK 9 000 CZK 12 000 CZK 10 000 CZK 8 000 CZK 25 000 CZK

RTB 2 10 000 CZK 108 000 CZK 33 000 CZK 69 000 CZK 57 000 CZK 38 000 CZK 148 000 CZK

... ... ... ... ... ... ... ...

... ... ... ... ... ... ... ...

... ... ... ... ... ... ... ...

SUM 1 300 000 CZK 1 300 000 CZK 1 300 000 CZK 1 300 000 CZK 1 300 000 CZK 1 300 000 CZK 1 300 000 CZK

Budget optimalization is an iterative process

budget shiftbudget shift

The optimal budget is reached when a channel reaches its maximum conversion.

Customer Journeys

RTB212636

RTB2157216

SEARCH

SKL2079081

RTB2125178

DIR

visualisation: Crossmasters

RTB and Display drive PPC and Search

conversion rate remained 24 %

CPA remained 0,019 CZK

2x more conversions

2,5x conversion value

Conclusion: last-click is a barrier of any growth

Data-driven attribution has sense with channels which shift customer in consumer funnel

Data-driven attribution gives immediate answers we couldn’t otherwise measure

High technology costs will return

The results are visible after some time (the need of enough data!)

Different marketing mix needs different model

scalability

all data at one place

ad-hoc reporting

transparency

At the end it’s a human job

“THE ONLY SOURCE OF KNOWLEDGE IS AN EXPERIENCE.”

ALBERT EINSTEIN

(1879 - 1955)