Intelligent Enterprise

Better Insight for Business Decisions

RSS
Webcasts
Whitepapers
Subscribe
Home

July 18, 2003

CRM: The Power of Prediction

By making smart use of data and information, predictive modeling and analytics can lead to vastly improved customer relationships — and for your organization, intelligent, cost-efficient sales and marketing

by Hussain Sabri

Continued from Page 1

You can see that some of these parameters, such as PE, are invaluable to the design process of products and services. Such a parameter "fact" plays a key role in validating promotion expenditures based on projected-to-actual deviation of results.

So, the road is paved to the sales force — or is it? The fact is that most list brokers don't offer customer-level information. Instead, they offer information (such as income or family size) amalgamated for a group of people (such as 9-digit or even 5-digit ZIP codes) along with their customer attributes.

Statistical Modeling To The Rescue

A clustering algorithm can split the demographics of customers compared to prospects into lightly overlapping groups. Clustering is the development of a predictive model that labels a new instance (in our case, a finite set of demographic attributes) as a member of a group of similar records (a cluster). The number of clusters can either be specified or detected. For marketing purposes, the practice is often to split the customer spectrum into a finite number of clusters to simplify the product packaging process.

Once we have concentrations of demographic attributes, we can apply a second technique, affinity modeling, to predict which products and services sell best together. In its simplest form, we perform affinity modeling by designing a "correlation coefficient" calculation formula. The correlation coefficient usually ranges between -1 and 1. It measures the degree to which we can relate two continuous columns. Usually, the coefficient is denoted by r, which measures the linear association between two variables. If a perfect linear relationship with a positive slope exists between the two variables (for example, between a bank account balance and accrued interest), we have a correlation coefficient of 1. Then, if this positive correlation exists, whenever one variable has a high value, so does the other. If a perfect linear relationship with a negative slope exists between the two variables (such as that between stocks and cash balance in a brokerage account), we have a correlation coefficient of -1. In general, if a negative correlation exists, whenever one variable has a high value, the other has a low value.

Facing Reality

By now, we have estimated what to sell to whom (and for that matter, when to sell it to them). Should we commence action and contact all of our clients and prospects?

Not yet. The world isn't a perfect place — and neither is any given products and services package. We need to estimate the probability of a particular customer acquiring a particular package. Therefore, the burden is on the data analyst to provide a prediction about what kind of customer fits the bill for a particular products and services package. Regression is the appropriate technique to achieve this pairing.

Unlike estimated results, the data that all regression algorithms use as input is real; it's data of past choices of packages of products and services by our segmented customers. The data tells us the accumulated, actual acquisitions (by customers and clients) of products and services offered by the company.

Understandably, we will discover some diversity in what we could interpret as a perfect match between customers and products. In Figure 2, a generic demographic attribute of the customer (represented in the horizontal axis, using income for example) produces a representation of the likelihood (represented in a vertical axis) that this customer will acquire the product. Obviously, the customer's preference varies: but for marketing, we need a baseline. This is what the red line in Figure 2 gives us. The baseline comes from a calculation based on data deemed acceptable (that is, the data inside the rectangle in Figure 2).

Producing that central regression line is pivotal to the whole process. Once you have it, it becomes the guiding basis for future product and service marketing to customers based on finite ranges in segment-defining attributes. (For example, number of children is an important demographic factor for life insurance packaging.) This baseline is most valuable when used outside the range of data you have — in other words, for predicting matches between customers and products and services. You can expect that the further you stray from your "confidence interval" (that is, the regression line range for which you have actual data), the wider the margin of error. You can see this effect in Figure 3. There, the portion of the regression line adjacent to the bottom-left corner of the acceptable data zone (outside the inner rectangle) isn't deemed to be of applicable significance: and therefore, you wouldn't expand the regression line for it.