Ronald serves as Director of Analytics at MeritDirect, the nation's leading business-to-business list brokerage and management company. He is responsible for the development of predictive models and provides research and analysis services to B-to-B direct marketers.
Prior to joining MeritDirect in 2001, Ron held senior level positions at BrandDirect Marketing, Bank of America and Mobil Oil Corporation. Past responsibilities have included the management of modeling and analysis teams, development of data warehouse applications, design of relational customer databases and administration of marketing information systems.He is a current member of the Direct Marketing Association Analytics Council and holds an MBA in Operations Research from St. John'’s University.
Here we go with the interview!
Question #1: How do you use math and statistics to identify behaviors that cause customers to respond to direct marketing activities?
Much of the analysis required to understand and prepare data for predictive modeling is facilitated by software that is readily available in the marketplace and can be customized over time. At Merit Direct, we have developed a library of programs, spreadsheets and analysis aids to facilitate quality built models. We apply a variety of SAS/STAT procedures and macros to gain a thorough understanding of potential predictors of response in what is referred to as data transformation. Transformation is where art meets science in the model building process and is believed to be the most important step in building predictive models. A thorough understanding of a business and how data is derived, acquired and applied can make a huge difference in the quality of a predictive model. Our analysis begins by examining the influence of every potential predictor in what is termed univariate analysis. Independently, each variable will increase, decrease or have a neutral impact on the likelihood of generating a response. The output of univariate analysis will establish a foundation of knowledge that can later be used to help confirm the results of a model and/or validate past and future selection strategies. A model, however, represents the way individual variables interact to have a combined influence on predicting response. This is referred to as the multivariate effect. An experienced modeler can review a list of variables qualifying in a model and identify those which provide a dominant influence, those which contribute the same or similar influence and those that could be eliminated without degrading the expected benefit of a model. The end result is a model that taps into only the most influential variables that work together to have the greatest influence on predicting response.
Question #2: Communication can be a challenge, especially between analytic individuals, and those who manage marketing, merchandising or creative departments. What does your team do to improve your ability to communicate with your clients?
Communication is a key factor to ensure a successful modeling project. At the outset of a project, it is essential to assemble all parties that have a vested interest in a model. This includes individuals who will secure data for the development extract, those who will apply the model, those who will assess the performance of a model and those who will benefit from the model. The first step towards a successful model depends on a consensus being formed on the objective a model or what type of behavior will be predicted. In general, at Merit Direct, we seek to predict response though it is important to define what type of response we are attempting to predict: first time response, repeat buyer, reactivation or high value response are a few examples. After setting the objective, we look at how the model will be incorporated into existing selection strategies, how the model will be tested and, finally, how results will be evaluated to validate the effectiveness of a model. As an analytical individual, I am mindful to communicate in simple business terms when discussing the development, expectation and application of models with end users and clients.
Question #3: What changes have you witnessed in the business intelligence and data mining profession over the past ten years? How have these changes impacted your job?
The largest change is the breadth and depth of data available through cooperative databases such as MeritBase. It is difficult to create effective models when more than half a dataset is devoid of critical information such as firmgraphics (SIC, Employee Size). And it is difficult to gain a complete measure of what purchasing power a prospect brings without the meta-data that reflects not just compiled data but some reflection of actual activity. Coops such as MeritBase can offer multi-sourced enhancement, behavioral meta-data, and of course a viable database to apply completed models and use them effectively without increasing list costs. Additionally, a number of analytical software packages have been introduced to the marketplace that streamline the process of evaluating data and developing predictive models. However, instead of expecting pre-packaged software to replace skilled analysts, it is more realistic to consider new software offerings as tools that can be used by individuals to gain greater intelligent insight into the data. While we can now do more with less and do it faster, the 'black-box' approach does not result in building stable, predictive models.
Question #4: Are there differences in how you approach the analysis of direct mail verses online? Do you observe differences in customer behavior, between these two channels?
Analysis of on-line responders is critical to a comprehensive measurement of ROI and a successful modeling project. Ignoring 1/3 or more of responders (white mail, web responders, etc.) can lead to a model skewed towards only easily tracked responders. Our studies have consistently shown that the hardest responders to track frequently come from large or institutional businesses and more often than not represent high lifetime value (LTV) customers. Conversely, web-responders tend to come disproportionately from single employee-home based business (SOHO) audience and may represent lower LTV customers, but are still essential to measure to accurately gauge performance of this large segment. At MeritDirect, we have successfully capitalized on capturing a history of past actions by compiling response over time and across all mailers who subscribe to our cooperative database. Additionally, we developed a proprietary matchback process, called MeritMatch, to gain the most comprehensive view of response across channels so that the modeler has accurate information to start the actual analytical processes.