Macy’s Inc: Five steps for getting more out of your customer data
Statistics chief sculpts customer data into a work of art
Every interaction with a customer is an opportunity to get useful data. Never before could marketers compile so much information about customers and markets, transform that information into useful knowledge, and guide the investment of resources with such precision.
Trouble is, data pours in from every conceivable channel, from incompatible computing platforms in mismatched data definitions and redundant customer entries. Transforming all this data into a clean, analysis-ready format can be a daunting challenge.
It's like creating a sculpture
When carving his masterpieces, Michelangelo felt that the figures were within the stone, and it was his job to liberate them. Similarly, it is our job to liberate the customer insight from all of the data.
The modeling process is a series of five steps – equivalent in some ways to the artistic process, whether working with stone, wood or data.
Know what you have been commissioned to sculpt.
Is it an enterprisewide promotion or a brand-building task? A one-time mailing or an integrated campaign? A one-day sale or a loyalty program? What is the target? You have to have a clear idea of what the analytics are intended to reveal, because the most appropriate modeling approach will depend on what you're trying to achieve.
Here is where the marketing community can make its most valuable contribution. There's a certain amount of information you can glean from models, and models provide a structure for decision making, but the business community has to engage and decide upon what will create success.
Choose the right tools for the job and the medium.
Are you going to press forward with intuition or take advantage of advanced analytics? Does the marketing team gather in a big room to brainstorm, or do you put yourselves in the hands of mathematical models? Both approaches have their merits and limitations.
An intuitive approach is easy to understand and communicate, because the logic will generally be self-evident. It is quick and easy to change course if something isn't working as expected. However, you could course-correct so many times that you lose track of the process. In contrast, an analytic approach provides a consistent and reproducible structure for the decision-making process. Everything has a place and everything makes sense. If your data went through all of the steps properly in the modeling procedure, you're going to get the results.
The best approach is to marry creative vision with technical expertise – the qualitative with the quantitative.
Find the right piece of marble or wood.
In data modeling, that means capturing the right customer universe and variables that will reveal your ideal target customers, working from raw material – data – that is as high-quality and free of defects as possible.
An interesting angle to the concept of data quality is that analytical models travel freely through time – learning from the past, projecting into the future, and using that speculated future to create information for the present. This time-traveling ability blurs the concept of "now" and can distort your data if you're not clear on how the time dimension affects the information.
Your data captures a past "now," which is a picture of what customers looked like in the past and how they responded to a previous event. You have a future "now" that describes what customers will look like on the day you want to select them. Then there is an even more far-reaching future "now" that describes how customers will likely look after some future event that is similar to the past event. For example, will they buy or not buy? How much will they spend? Eventually, all of these "nows" will become past "nows," and the model moves on to looking at future "nows" to influence a new present "now."
Sound confusing? It can be. If you're not careful, this time traveling can cause damage to your data. We call them temporal leaks, where you inadvertently leak information from the future to the past. Data sets built during the modeling process incorporate forward-looking insight about events and behaviors that haven't actually happened. To preserve the quality of your raw material – the data foundation for modeling – the process must not taint the data by using future unknowns in the same way as past certainties.
Chip away at everything that imprisons the figure.
Chip away all the extraneous data and variables that don't matter. We have lots of numbers in front of us, but only certain ones are actually useful. The process entails data wrangling to make the data model-ready, then rough-cutting the data to remove extraneous data points and variables.
We might initially start with 1,000 variables and reduce that to about 100. Think about hearing somebody sing in a chorus of 1,000 people. After the performance, a singer may ask how he or she sounded. You say, "You were wonderful," but you really have no idea. If you start to break down your chorus into smaller and smaller ensembles, you can begin to hear individual voices more clearly. That's what the data reduction process achieves. It creates an environment where the true voices become clear.
Polish and present your creation.
Build, test, refine and execute analytical models that will deliver on your desired goals. You're not limited to just one type of modeling method. Some are more insightful, others more mathematical. Not all give the best results for all applications. The idea is to choose the modeling routine that works the best for your application. That means the model assessment stage is quite important.
Multiple models can work in complementary ways to create a big picture. At Macy's we build at least two models for most events: one that predicts the probability of someone shopping, and another that predicts how much that person is going to spend. We combine these models to maximize the analysis.
For a high-volume marketer, even a small increase in campaign response rates can make a significant impact on the bottom line. Macy's use of analytics over the last two years has improved marketing results by millions of dollars. And that's just for direct mail campaigns.
Bio: Paul Coleman, PhD, is the Director of Marketing Statistics for Macy's Inc.
Paul Coleman, Macy's Inc.