Using Predictive Models, Part 1

  |  May 17, 2005   |  Comments

A predictive modeling primer for marketers. Part one of a series.

Why will marketers spend literally hours discussing whether the a headline should be 12- or 16-point type but completely neglect to use analytics to investigate their consumer data and increase response rates?

In a recent survey, I discovered many marketers are interested in more information about predictive modeling. Their self-reported interest level is high, but their knowledge level is low.

If you're a marketer, especially in the telecommunications, financial, insurance, credit, banking, or retail industries, and you don't extensively use sophisticated predictive models to increase marketing return on investment (ROI), you need to read this column and implement predictive modeling.

Part one explains the basics of what a model is and how it's developed. The following parts will describe different types of algorithms (regression, CHAID, etc.) used in model development, as well as how models are used.

What Is a Predictive Model?

A predictive model is simply an equation used to predict something.

In a silly but effective example, I often illustrate what a predictive model is by collecting shoe sizes and heights from a group of people. Then, using SPSS (define), I make a quick x/y plot that shows how generally taller people have larger shoes sizes than shorter people. Then I use SPSS to quickly determine an equation that allows me to predict someone's shoe size if I know his height.

That equation is a predictive model. That's really all there is to it.

Obviously in marketing, you're not usually interested in predicting a customer's shoe size, but you are interested in predicting:

  • How much money a customer will spend on your products

  • How likely someone is to respond to an offer

  • How likely someone is to stop using your services

If the correct data are available, we can use them (along with some sort of statistical software package) to develop a predictive model that will tell us the answer to each of the above questions. Marketers then use the predictive model (i.e., the equation) to score and rank prospects based how much money they're likely to spend, how likely they are to respond to an offer, and so on. Marketers can then tailor their efforts accordingly.

Why Is It Important?

Direct marketing's most important aspect is your specific target market. In the case of email, direct mail, or outbound telemarketing, your target market is specified by your lists. It doesn't matter how good your product or service is or how wonderful your creative. If you send your piece to the wrong list, results will be poor.

A predictive model provides you with the best names possible. That results in the best possible response to your marketing efforts.

How Are They Developed?

To develop a predictive model, you must first decide exactly what it is you want to predict. There are many possibilities, including:

  • Who is most likely to respond to a mailing?

  • Who is most likely to buy a specific new product or service?

  • Who is likely to defect to a competitor?

  • Who is likely to be a high credit risk?

  • Who is likely to be a high-value customer?

After you know what you're trying to predict, you then must make sure the correct data are available for model development. If you don't have appropriate initial data from which to build the model, a sample test may have to be run to collect some. You might conduct a small test campaign for a new product across a number of market segments and use the resulting data to develop a model of who buys the product.

Once the required data set is acquired, statistical analysis software (e.g., SAS (define), SPSS) is used to determine which variables influence what you're trying to predict. You might find, for example, the response rate to your mailing varies by recipient's income and age. Income and age are then considered "predictive" variables, since you can use them to predict how likely someone is to respond. After determining which variables are predictive, the analysis software determines the "weights" (or "coefficients") for each predictive variable.

Various Algorithms and Software

There are a few different algorithms from which an analyst or marketer can chose when she wants to develop a model. I'll discuss those in part two (as well as the software packages that feature the algorithms) and describe how models are used. I'll also discuss the results they can help you achieve.

ClickZ Live Toronto On the heels of a fantastic event in New York City, ClickZ Live is taking the fun and learning to Toronto, June 23-25. With over 15 years' experience delivering industry-leading events, ClickZ Live offers an action-packed, educationally-focused agenda covering all aspects of digital marketing. Register today!

ClickZ Live San Francisco Want to learn more? Join us at ClickZ Live San Francisco, Aug 10-12!
Educating marketers for over 15 years, ClickZ Live brings together industry thought leaders from the largest brands and agencies to deliver the most advanced, educational digital marketing agenda. Register today and save $500!

ABOUT THE AUTHOR

Brian Teasley Brian Teasley is the leader of Teasley, a consultancy that helpsadvertisers, marketers and advertising agencies use data and analysis toimprove their marketing campaigns. Brian has over 14 years experience inengineering and marketing, and has worked for numerous Fortune 100companies. Brian also teaches a marketing course at New York University. Heholds a M.S. degree in Applied Statistics from Iowa State University and aBA in Mathematics and a BA in Mathematics and Statistics from St. OlafCollege.

COMMENTSCommenting policy

comments powered by Disqus

Get the ClickZ Analytics newsletter delivered to you. Subscribe today!

COMMENTS

UPCOMING EVENTS

Featured White Papers

Gartner Magic Quadrant for Digital Commerce

Gartner Magic Quadrant for Digital Commerce
This Magic Quadrant examines leading digital commerce platforms that enable organizations to build digital commerce sites. These commerce platforms facilitate purchasing transactions over the Web, and support the creation and continuing development of an online relationship with a consumer.

Paid Search in the Mobile Era

Paid Search in the Mobile Era
Google reports that paid search ads are currently driving 40+ million calls per month. Cost per click is increasing, paid search budgets are growing, and mobile continues to dominate. It's time to revamp old search strategies, reimagine stale best practices, and add new layers data to your analytics.

WEBINARS

Resources

Jobs

    • GREAT Campaign Project Coordinator
      GREAT Campaign Project Coordinator (British Consulate-General, New York) - New YorkThe GREAT Britain Campaign is seeking an energetic and creative...
    • Paid Search Senior Account Manager
      Paid Search Senior Account Manager (Hanapin Marketing) - BloomingtonHanapin Marketing is hiring a strategic Paid Search Senior Account Manager...
    • Paid Search Account Manager
      Paid Search Account Manager (Hanapin Marketing) - BloomingtonHanapin Marketing is hiring an experienced Paid Search Account Manager to...