Help & Support

212.660.6600

5 Leading Enrichment Datasets - How Do They Compare?

In order to improve direct mail response rates, we have selected the top datasets to enrich our target audience data. By enriching the data we develop a fuller picture of each audience member, one that allows us to more accurately target only those individuals most likely to respond. So, if a 10,000 piece mail campaign usually receives 100 responses and our data enrichment and targeted segment selection increases that number to 120, 160 or even 280, then we will have contributed to a significant boost in the campaign’s profit-margin.    

It’s normal to have some information about potential customers, such as their Name, Place of residence, Email, Facebook account, Twitter account, etc. While this information has some token value, it falls short of providing real value for any sort of sophisticated analysis. This is especially true when the information is processed through artificial intelligence algorithms.

Many additional data characteristics are available for members of our marketing audiences, including their Travel, Sports, Book, Nutrition, Music, and Hobby preferences. These features can be combined with Marriage status, Religion, Gender, Age, Spoken language and more.

Using machine learning algorithms, we combine our original data with these new datasets to identify customer behavior patterns, giving us a much richer picture of our targets. With this data, based on previous behavior, we identify which people are most likely to respond to specific campaigns.

Five Data Sources

We examined a group of five different data sources, each providing individualized datasets. 

We started with a dataset which includes Name, Location, and Target (as in, whether a person bought a product or not). Each data source delivered slightly different information in their datasets.  

Here is a brief overview comparing the enrichment options provided to us by some of these sources.

Source A has two data products, (labeled A, and A2 respectively) both of which include gender, amount of money spent and purchasing interests.

Source A

Data originates from:

  • Public records — including phone books, voter registration files, deeds and permits

  • Surveys — a collection of self-reported data from 20 million households in North America

  • Partners — including magazine publishers, catalog purchasers, B2B sources and more

Source A2 

Includes multi-sourced transactional data with data on market channels and purchases.

Source B

The Source B dataset includes over 1 billion records with 2,200 features. The dataset includes Email, Phone, Postal address, IP addresses, Website cookies, and more.   Information is updated daily.

Selects provided include a wide range of categories, including:

  • Geographic data

  • Demographic data

  • Psychographic data

  • Ethnicity

  • Various interest categories

  • Political

  • Real-time sales leads based on the following categories

    • Home improvement

    • Automotive

    • Solar

    • Insurance

    • Payday loans

Source C

Source C contains address and telephone data.

The Source C dataset includes 120 million consumer, business, and government records, including consumer telephone data with area code correction, and business telephone data.

There is also an address database with 160 million records, including records of those who moved in the past 4 years, zip+4, carrier route numbers, verification of building/firm name, business vs. residential indicators, rental vs. ownership, seasonal information, occupation and vacancy information. 

Beyond this there are 200 million user records with demographic and lifestyle selects, including financial and purchasing behavior.

Source D

Source D includes data from 240 million consumers in more than 140 million US households. They include donation information from approximately 116 million people and data from 180 million mail order purchasers. The dataset includes 115 million landline and 65 million cell phone numbers. The selects contain demographic and lifestyle attributes, as well as information about life changes.

As you can see, there is considerable data overlap among the different data sources. However, the overlap is clearly not 100%, and the results show different degrees of predictability of consumer behavior.

Methodology

We examine the quality of data provided to determine how we could best increase response rate. We often improve the data via steps like calculated features. For example, if a client has physical locations, the distance from a location might be a strong indicator of likelihood to buy.

By using a series of algorithms, we attempt to calculate the “lift” value — a measurement of how many times better our model would perform over a random sample — of certain patterns within our data. In order to do this, we combine two datasets and then perform the following tasks:

  1. Score all participants

  2. Use these scores to target the groups most likely to respond

  3. Focus the campaign on those who score the highest

We then look at the models we have created and determine the lift factor.

If our calculations are correct, higher scores should result in more purchases. The better the model, the larger the difference between the scores of the responders and non-responders (likely non-responders should get a score of less than 1.0). Using this information, we can identify those aspects within our datasets that are likely to have the greatest amount of positive impact to the bottom line of our campaigns. We are able to get a clear picture of whether a target attribute of our subjects has a higher likelihood of getting a response.

The Results

We ran a series of tests against identical marketing campaign data and were able to identify some variability between the effectiveness of models provided by each dataset.

For instance, here is a chart of what occurred with the first campaign, labeled APP01.

 

Campaign ID

Experiment name

Dataset

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

APP01

17MKFL10

A2

1.45

1.33

1.27

1.23

1.18

1.14

1.09

1.06

1.03

1

APP01

17MKFL10

B

1.68

1.49

1.4

1.32

1.26

1.21

1.17

1.12

1.06

1

APP01

17MKFL10

D

1.76

1.59

1.45

1.37

1.29

1.23

1.18

1.12

1.06

1

APP01

17MKFL10

C

2.92

2.21

1.86

1.64

1.5

1.38

1.29

1.2

1.1

1

APP01

17MKFL10

A

1.99

1.77

1.6

1.49

1.4

1.31

1.24

1.16

1.08

1

APP01

17MKFL10

A reduced 

1.45

1.41

1.37

1.35

1.28

1.24

1.19

1.19

1.08

1

APP01

17MKFL10

B (full)

1.69

1.53

1.42

1.34

1.27

1.23

1.17

1.13

1.07

1

APP01

17MKFL10

B (cleaned)

1.69

1.53

1.41

1.34

1.28

1.23

1.18

1.12

1.07

1

 

In this campaign we can see that dataset C’s model came out significantly stronger than the others. If we look at using, for example, 40% of the data, the model for dataset C performed 1.64 times better than using a random sample, for a lift of 1.64. The next best model was dataset A, which was able to return a lift of 1.49. The weakest response was from dataset A2 which returned a lift of only 1.23.    

In the second campaign, APP02, we saw slightly different results. However, dataset C’s model still produced the strongest results.

 

Campaign ID

Experiment name

Dataset

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

APP02

18MKFL10

B

1.62

1.43

1.35

1.28

1.23

1.19

1.15

1.11

1.06

1

APP02

18MKFL10

D

1.76

1.56

1.46

1.37

1.3

1.25

1.19

1.13

1.07

1

APP02

18MKFL10

A2

1.45

1.33

1.27

1.22

1.18

1.14

1.1

1.08

1.04

1

APP02

18MKFL10

C

1.99

1.69

1.53

1.4

1.32

1.24

1.18

1.12

1.07

1

APP02

18MKFL10

B (cleaned)

1.61

1.45

1.35

1.29

1.24

1.19

1.16

1.11

1.06

1

 

For example, looking at the 40% sample, we can see we were able to get a 1.4 times lift in response rate, compared with the next highest, which was in this case dataset D, which came in at 1.47. The difference here is marginal, but significant. When using smaller samples, the differences become more pronounced.

In the third campaign, dataset D came in considerably stronger than any of the other competitors.   

 

Campaign ID

Experiment name

Data source

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

APP03

A12A18

B

2.37

1.95

1.72

1.48

1.36

1.26

1.23

1.13

1.08

1

APP03

A12A18

D

3

2.37

2.06

1.76

1.58

1.43

1.33

1.2

1.1

1

APP03

A12A18

C

1.78

1.67

1.58

1.4

1.32

1.21

1.16

1.16

1.07

1

APP03

A12A18

A

2.97

2.26

1.87

1.55

1.48

1.38

1.26

1.19

1.11

1

APP03

A12A18

B(full)

2.33

1.83

1.67

1.47

1.31

1.23

1.18

1.11

1.04

1

APP03

A12A18

B(cleaned)

2.72

2.3

1.8

1.53

1.38

1.27

1.19

1.1

1.05

1

If we again look at 40%, we see a 1.76 increase in lift. At 30%, the rate of 2.06 means the return is more than double that of a generic campaign. In this case dataset A came in second, but considerably lower, at 1.55 and 1.87 respectively.

Summary

Choosing a dataset for marketing data enrichment can be a complicated process that requires experience. While there are factors that may make one dataset look more appealing than others, such as size, or richness of data, these don’t always translate into results for campaigns. Cost has some influence but not always. Sometimes a lower cost dataset can produce equally great results. Other times, spending more on data returns profits that exceed the extra cost. We did not include cost calculations above, but there is significant variability in the costs between different datasets.

At Wyzoo, by doing this sort of advanced analysis we are able to look more closely at the impact each dataset will have. We help make decisions regarding the effectiveness of direct marketing campaigns, and help determine the best course of action for targeting likely responders. We also get an idea as to the cost-effectiveness of using different and sometimes multiple data sources for enrichment. Results will always vary, as humans are among the most difficult datasets to measure. However, by increasing the likelihood that a person will respond to a mailing, we can have a large impact on the campaign’s bottom line.

Resources

Other Articles

ETL Tool Comparison: KNIME vs Alteryx

ETL Tool Comparison:...

Evaluation: Driverless AI vs. DataRobot

Evaluation: Driverle...

Choosing the Right Tools for Data Ingestion, Part 2

Choosing the Right T...

Evaluation: Auger.AI vs. H2O Machine Learning Tools

Evaluation: Auger.AI...

Related Experts

Data Analyst

Data Analyst

Data Scientist

Data Scientist

Machine Learning Engineer

Machine Learning Engineer
robo happy