Dept Agency16 November 2017

Build real-time Audiences based on Machine Learning

Thanks for having us!

Mustafa Himdi

Digital Strategist

Job Deibel

Data Scientist

Predictive Retargeting

Current retargeting practices

Standard Retargeting Audiences

All visitors Standard retargeting

Assumption

Assumption

“A site visitor is interested in my product/service”

Observation

“Job is a site visitor”

Conclusion

“Job is interested”

[PLAATJE WINKEL]

Custom Retargeting Audiences

Business knowledge

Analysis

Gut feeling

This hopefully results in higher conversion rateaverage

Analysis

Standard retargeting

Custom retargeting

1% conversion rate

5% conversion rate

5%3%8%6%2%10%40%30%10%15%25%3%1%1%

Analysis

average

Standard retargeting

Custom retargeting

1% conversion rate

5% conversion rate

Analysis

average

5%

3%

8%

6%

2%3%

1% 1%Standard retargeting

Custom retargeting

1%

5%80%

60%

55%

75%

95%

50%

Analysis

5%

3%

8%

6%

2%

80%

60%

55%

75%

95%

50%

3%

1% 1%

But we only want to target "high potentials”

Standard retargeting

Custom retargeting

1%

5%

"Half the money I spend on advertising is wasted, the trouble is I don't know which half"- John Wanamaker, 1908

How do we predict conversion intent?

Machine Learning

Algoritms Applied to

big data sets

Predictive

model

70%

Predict

purchase

probability

OK,But how does this work?

Data collection

Data source -User level

clickstream data

Data collection

Data source -User level

clickstream data

R-programming

Data collection

Data source -User level

clickstream data

Raw dataset -Training data

R-programming

Data collection Algorithm

Raw dataset -Training data

Machine learning

algorithms Neural Network

Naive Bayes Random Forest

Gradient Boosting Support Vector Machine

Classification Trees

Data collection Algorithm

Raw dataset -Training data

Machine learning

algorithms Neural Network

Naive Bayes Random Forest

Gradient Boosting Support Vector Machine

Classification Trees

Data collection Algorithm

Predictive

Model

Predictive Model

Raw dataset -Training data

Deployment

Data collection Algorithm

Predictive

Model

Predictive ModelData collection

Deployment

Data collection Algorithm

Predictive

Model

Predictive ModelData collection

JavaScript

in GTM

Deployment

Data collection Algorithm Predictive Model PredictingData collection

35%

75%

JavaScript

in GTM

Real-time

predictions

Deployment

Data collection Algorithm Predictive Model PredictingData collection

35%

75%

Real-time

predictions

Audience building & bid-strategy

Campaigning

Marketing Automation

First results

Facebook Retargeting1.

Facebook Retargeting - Audience1.

Audience Size 18640185Total Conversions (last-click)

Facebook Retargeting - Audience1.

High Conversion IntentLow Conversion Intent

Facebook Retargeting - Audience1.

Audience Size 1258371

Conversion Rate 0.56%Total Conversions (last-click)

Audience Size 6057114

Conversion Rate 1.88%Total Conversions (last-click)

High Conversion IntentLow Conversion Intent

48%

Audience Size 1258371

Conversion Rate 0.56%

Audience Size 6057114

Conversion Rate 1.88%Total Conversions (last-click) Total Conversions (last-click)

High Conversion IntentLow Conversion Intent

161%

Audience Size 1258371

Conversion Rate 0.56%

Audience Size 6057114

Conversion Rate 1.88%Total Conversions (last-click) Total Conversions (last-click)

High Conversion IntentLow Conversion Intent

Uplift Conversion Rate

236%

Audience Size 1258371

Conversion Rate 0.56%

Audience Size 6057114

Conversion Rate 1.88%Total Conversions (last-click) Total Conversions (last-click)

High Conversion IntentLow Conversion Intent

Continuous learning

What if…

20,000 website visitors

Strategy

High

Medium

Low

Testing

High

Follow up (50%)

No follow up (50%)

Testing

High

High frequency cap (50%)

Low frequency cap (50%)

What else?Hard CTA

(50%)

No discount (50%)

Low bid (50%)

High bid (50%)

Soft CTA (50%)

Discount (50%)

High

Alternative Applications

Alternative Applications

Activate dynamic content

Predict which channel to use for retargeting

Predict expected revenue when retargeted

Predict the ideal frequency cap

Make predictions on product category level

Send predictions to external datasource

Conclusion

All business problems have a data solution!

Any questions? Fire away.

Mustafa Himdi

mustafa.himdi@deptagency.com

Job Deibel

job.deibel@deptagency.com

The heart and mind of your digital business