We transform data into actionable information. Whether it is a program aimed at acquiring new customers or a program designed to build loyalty among an existing base, we can microtarget specific actions to specific individuals and make these programs more effective.
Data mining finds patterns in data. When these patterns are associated with a behavior of interest they can help you understand the causes of that behavior, or help you identify other individuals with similar behaviors.
Data mining usually involves five steps:
- Selecting the data you wish to study
- Transforming the data into a form that enables analysis
- Extracting information using statistical algorithms
- Synthesizing the information into an actionable form
- Communicating the findings in a clear and useful manner
We bring expertise in a broad array of statistical algorithms, including neural networks, genetic algorithms, decision trees (CHAID, CART, etc.), clustering algorithms, and logistic and linear regression. While a software vendor may often see every problem as a nail for his specific hammer, we approach each problem with an open mind. Every algorithm brings a unique set of advantages and disadvantages. We often apply multiple techniques to the same problem. But in the end, the algorithm is rarely the key to success. Success is usually determined by how well the problem is defined and formulated.
A predictive model is a mathematical formula that predicts individual behavior based on data. Once built, these models can help you choose individuals who are likely to have the behavior in which you are interested. The type of behavior will depend on your objective, but some of the behaviors we have predicted include:
- Responsiveness to specific offers or channels (mail, phone, email, etc.)
- Likelihood of using a specific product
- Interest in a candidate, product, or issue
- Likelihood of customer defection
- Risk of non-payment
The model development process is a complex combination of art and science. Successful modeling comes from correctly formulating the problem, understanding the data and choosing the appropriate statistical algorithm.
Segmentation is essentially a sorting task. The objective is to divide people into groups with similar behaviors or interests.
Segmentation gives you the ability to customize your communication for different groups. For example, in politics, the concerns of young married couples with children may be very different than the concerns of a widow on social security. Similarly, the financial needs of a young couple starting a family are very different then the needs of empty nesters approaching retirement. By identifying groups with similar needs and tailoring your communications appropriately, you have the opportunity to create a relevant, compelling conversation.
How does segmentation work?
The best statistical model is of little value if it is used incorrectly. The value is created when all of the pieces are optimally integrated. We can help you deploy the use of models and segmentation into your day-to-day operations to increase your efficiency and effectiveness.
We begin by working with you to define clear, measurable objectives. For a marketing campaign, the measure of success can range from simple response rates to lifetime customer value. For a political campaign, success can be measured in vote movement, a change in a favorability rating or name recognition. We use experimentation to identify the actions that perform the best. If the ultimate goal is a long way off, it is often necessary to develop proxies that allow us to measure intermediate progress. While some measures provide immediate feedback, others look at the impact on a customer’s lifetime value. If your organization does not have a measure of customer value, we can help you define one. If the relevent measure is not immediately available, we can develop proxies that allow us to measure intermediate progress.
We encourage our clients to learn through testing. A well-designed campaign can measure impact and identify opportunities for improvement. When appropriate, we are happy to work with your vendors to make sure programs are executed correctly so that robust learning can take place.
Optimization Overview (pdf)
Customer-value should take into account the cost of acquisition, operations, retention and risk, as well as observed and potential revenue.
We use testing to measure the value of one set of actions as compared to a different set of actions.
A good test uses an experimental design that maximizes learning while minimizing the cost. There are two characteristics of a good experimental design. First, the design allows you to distinguish the effects of an action from the effect of being chosen to receive that action. For example, if you want to compare the impact of two different pieces of mail, you need to make sure the two groups receiving the two different mail pieces are identical in all other ways. Second, the sample sizes need to be sufficiently large. The experimental design should generate results that can measure the outcome accurately and with statistical confidence.
Since the overall effectiveness of a given test can be obscured by conflicting behaviors among sub-groups, it is necessary to analyze test results at multiple levels. For example, if one group responds negatively to a mail piece, while a second group responds positively to that same piece, the overall results may look like the mail is having no effect. We look at impact by segments to make sure that any effective treatments are not overlooked.
Finally, the experimental design determines how cost effective the test is. We can implement sophisticated factorial design techniques that allow us to test many things simultaneously and still get statistically valid measures while restraining sample size.
The value of a given model is a function of how it is built and how it is deployed. We can assess your current model development process to ensure the best-in-class procedures and methods are used. We also examine how those same models are deployed to verify that their use is consistent with how they were built. As part of this review, we look for interactions between models, or between models and operational processes.
We bring deep understanding of all stages of the data life-cycle, from its collection, organization and storage to its transformation and use. Our experience spans from the design and development of large data warehouses to the daily use of their content. Our objective is to identify how the data life-cycle can be changed to improve the quality, ease-of-access and usefulness of the data. We identify gaps and recommend improvements.
Read on to discover our approach to developing data strategies.