Find Model Customers with Predictive Analytics

W W W . R E L E V AT E G R O U P. C O M

6 Steps to Find Model Customers

G O F R O M TA R G E T I N G P E O P L E W H O M I G H T R E S P O N D T O F I N D I N G P E O P L E W H O W I L L R E S P O N D .

It’s your job to know your customers, what makes them buy and how to find more like them. The hard part of that job has always been separating the wheat from the chaff.

But if your marketing budget is being tracked to sales and ROI more than ever, you just can’t afford for half of it to be wasted. And even in media that aren’t constrained by budget (like email) there’s a price to be paid for wasting marketing touches (like being labeled a spammer).

6 S T E P S T O F I N D M O D E L C U S T O M E R S G O F R O M TA R G E T I N G P E O P L E W H O M I G H T R E S P O N D T O F I N D I N G P E O P L E W H O W I L L R E S P O N D


“I KNOW HALF THE MONEY I SPEND ON ADVERTISING IS WASTED. I JUST DON’T KNOW WHICH HALF.”

—John Wannamaker

It’s more important than ever to be able to target your marketing efforts not just to people who might be interested, but to the people who are most likely to be interested.

The best way to do that is to use past performance to predict future behavior, and data modeling with predictive analytics is the key to that. Data modeling takes into account the interaction of data elements that, in combination, allow you to identify the people on a list who are most likely to take the desired action.

So stop targeting the names you happen to get, and use these techniques to target people who’ll be your model customers.

120%

100%

80%

60%

40%

20%

0%

0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50% 55% 60% 65% 70% 75% 80% 85% 90% 95% 100%

Random

Wizard

Validation

Estimation

% OF INITIAL POPULATION

% O

F TH

E M

AX

EXPE

CTED

PRO

FIT



How Data Models HelpData models can be separated into two main types and strategies: Customer Models analyze the behavior of people who have already done business with you, and Acquisition Models help you identify prospects most likely to respond to your offers.

In each case, the idea is to sort the list by significant variables so you can contact a smaller subset of it to get more response. In the Lift Curve table below, the “Random” line represents your typical campaign, where each 10% of the contacts made bring in roughly 10% of the total response—so to get 80% of the responses, you need to mail 80% of the list. The “Wizard” line is an ideal world where you could get 100% of the response by mailing just 10% of the list (essentially mailing only the people you know will convert). In the real world, no model is going to let you get that level of response, but it helps to illustrate the idea.

The “Validation” and “Estimating” lines represent what data modeling lets you do. These are the results of modeling in the real world from cases Relevate did for its clients. The “Estimation” line shows the results the data model predicted, the “Validation” line shows how that model would have

performed based on those individuals’ reactions to previous campaigns.

In short, the model shows how you can get 80% of the response with only 55% of the names, and the data validation shows that those gains were real and repeatable for the clients.

While those are simple concepts, they can be applied in many ways to ensure successful campaigns. For example, customer models can be very valuable when used to identify current customers who will respond best to campaigns built to optimize:

• Retention

• Reactivation

• Cross-Sell

• Lifetime Value

Retention is an essential aspect of marketing, but you downloaded this whitepaper to learn how to find new model customers, which is where acquisition models come in. Used properly, acquisition models can increase response rates in your efforts to generate new customers. This means your prospecting can be conducted more profitably, and your prospecting budget will yield more new customers.

Step 1 - Data Input

Step 2 - Data Prep

Step 3 - Model Data

Step 4 - Extract Samples

Step 6 - Validate Entire Data Set

Step 5a - Model Creation Step 5b - Model Validation



Acquisition ModelsThere are generally two types of data models you can use to identify better customers. You can look at how prospects have responded to other marketing campaigns, or you can look at your existing “good” customers and use them to build a model that will help identify more customers like them.

In each case, you’re creating the model to find characteristics responders have in common that you can use to refine prospect lists and target only the segments that share those characteristics.

• Response Model

◦ Data set: Analyze a sample of solicitations and responses from prior campaigns.

◦ Action: Identify variables that differ between those who took the action (response) vs. those who did not.

• Good Customer Match Model

◦ Data set: Analyze a sample of your best customers.

◦ Action: Compare that to a sample of your list that have not yet taken the buying action to identify variables that set them apart.

Relevate’s 6 Steps to Building a Great Acquisition Model

At Relevate, we create these models for many of our clients, and we’ve boiled the process down to six essential steps. You can use these steps to build your own acquisition model, or contact a data professional who can create it for you.

1. Gather the data set(s): If you’re building a response model, this means pulling together the solicitation and response data to be modeled. If you’re building a “Good Customers” model, identify those customers and compile the relevant data.

2. Standardize the data: Often this data comes from a variety of sources, and will need some work to prepare it for enhancement.

3. Append more data to that data: The modeling data set is then appended to the total universe provided to add more information that helps delineate how these customers and prospects behave. The data Relevate appends to these models includes hundreds of variables identifying consumer demographics (census data, lifestyle data, buying tendencies, etc.) and business firmographics (number of employees, sales volume, ethnicity code, woman-owned, Fortune rank, etc.) that allow you to build a 360-degree view of your customers.

4. Extraction: From the appended data set we extract two samples. We use the first to create the model, and the second to test the model.

5. Create the model: We typically will run through the modeling process a few times until we get a combination of variables that produce a scorecard that is both statistically strong and robust (more on this below). The resulting scorecard may only have 15 to 25 variables, but those are the variables that, when combined, will provide the best lift. Those are the variables we’ll use to sort prospect lists to increase response rates. In the Lift Curve chart above, this was the “Estimation” line.

6. Validation: Finally, we apply the new model to the entire data set that was provided. In order to see what would have happened had we used the model in the past, we “back cast” it against past actual results for the model data set. This allows us to verify that the behavior we’ve modeled syncs up with real-world results. In the Lift Curve chart above, this was the “Validation” line. When the validation line and estimation line are very close, as they are in the chart, that means the model is on target and repeatable.

Using the Models

Going back to the case examined in the Lift Curve chart above, here are the 15 variables the model identified as key contributors. The variables include some demographics and some buying behavior indicators. In this model, AGE has the highest contribution and Year of Vehicle has the lowest contribution.

When scoring a data set, each individual will be assigned a point-value for each one of these variables to create a total score. The total scores can then be ranked from high propensity to respond to low propensity, and can be broken out into segments, such as 10% increments (“deciles”) based on score.



Variable Contributions ChartResponder Model - Variable Contributions

18%

16%

14%

12%

10%

8%

6%

4%

2%

0%

Age Ethnicity TotalOnline $

House-Hold

Income

VehicleType

DwellingType

InsuranceResponder

HomeOwnership

ContinuityProducts

VehicleYear

RetailTotal $

CatalogBuyers

NetWorth

DirectRepsonse

Orders

Here’s what the scorecard looks like for the data used in the Lift Curve chart above.



ConclusionJohn Wannamaker may not have known which half of his advertising wouldn’t work, but with data modeling

and predictive analytics, you can get a pretty good idea of exactly that for your own marketing.

By eliminating names from the list that are less likely to respond and zeroing-in on model customers, your prospecting becomes much more efficient. If you’re looking for a way to lift response rates and eliminate

waste from your budget, ask your data scientist to run the models and cut to the top of the list.

The top four deciles individually produce index levels greater than 100%. In this case, we would recommend the client select records in the top one or two deciles in an initial test. We would also recommend a random sample among all deciles, just to validate results in a live solicitation.

If the test verifies the model is working, this list can be solicited at a much higher rate of return than unmodeled data, because you can sort it to zero-in on just those customers who are more likely to respond—either because their response patterns dictate higher response, or because they already fit the model of your best customers.

1

2

3

4

5

6

7

8

9

10

150,928

157,679

159,319

165,442

167,495

175,258

171,858

176,619

181,325

182,226

9%

9%

9%

10%

10%

10%

10%

10%

11%

11%

1,618

1,139

981

844

721

585

511

401

311

176

22.2%

15.6%

13.5%

11.6%

9.9%

8.0%

7.0%

5.5%

4.3%

2.4%

1,618

2,757

3,738

4,582

5,303

5,888

6,399

6,800

7,111

7,287

22.2%

37.8%

51.3%

62.9%

72.8%

80.8%

87.8%

93.3%

97.6%

100.0%

Decile Total Customers Responders

QTY %TTL QTY %TTL Index Cumulative QTY

248

167

143

118

100

77

69

53

40

22

Cumulative %TTL

Total 1,688,149 100% 7,287 100% 100

Total File “Gains” ReportTotal Sample Scored by the New Response Model Data Set

GRAND TOTAL 2,581,891 11,924

Unscored 893,742 4,637

Data & Analytics

Find Model Customers with Predictive Analytics