The Drill Down

How your operations and marketing efforts can benefit from statistical analysis and modeling.

Forgive me if I generalize for a minute. There are two approaches to marketing analysis: the arithmetic and the statistical.

The Arithmetic Approach

Sometimes called “descriptive analytics,” this is relatively straightforward and inexpensive, depending on a spreadsheet and the sweat of your brow. Extracting a season’s sales from your transaction system to your spreadsheet, you can determine the following:
- percent response, by dividing your number of orders by your mail quantity per segment;
- average order value, by dividing your gross sales by your number of orders per segment;
- contribution per segment, by subtracting your cost of goods, promotional costs, and offer costs from your gross sales;
- contribution per order, by dividing the contribution per segment by its number of orders; and
- dollars per book, by dividing gross sales by the number of pieces mailed to each customer segment.

Of course, you have to use a best-guess formula to account for unsourced Web orders, but such formulas can be more than adequate to decide on next season’s mailings and promotions. Moreover, the arithmetic method also can yield reasonably good segmentation analysis, based on pre-season and post-season analysis of recency, frequency and monetary (RFM) value.

Statistical ROI

Why, then, would you want to invest in the other approach — statistical (sometimes called a “predictive” approach) — to marketing analysis? The benefit here is two-fold:

1. Because the first step in statistical analysis is creating a normalized data mart or data warehouse, your data will be cleansed of anomalies, duplicates and missing pieces.

2. And because the statistical approach is predictive, you can make more precise or more granular decisions about likely future customer behavior.

The statistical approach also will help you to:
- determine the likelihood a customer will become a repeat buyer and that a repeat buyer will evolve into a high lifetime value customer;
- improve promotion response rates by differentiating responders from non-responders;
- identify product cross-sell and upsell buying propensities.

Most important of all, the statistical approach is based on modeling, which allows you to test the validity of your projections before acting on them. The larger your housefile, the more numerous your promotions. And the more complex your merchandise, the more likely you are to benefit from the modeling approach.

Case Studies in Modeling

One of the largest gardening supplies catalogs found its RFM models were inadequate when it acquired a competitive catalog a few years ago. Concerned with potentially wasting many catalogs on unresponsive customers, missing revenue opportunities and experiencing higher costs than necessary, it decided to take a different modeling approach. Its executives especially wanted to improve their method for distinguishing avid gardeners from casual buyers to determine which catalog each customer would be mailed.

Investing in a statistical package, the models developed helped company officials classify more than 85 percent of customers, compared to just 50 percent using their previous methodology. Putting their new analytical approach to the test, response rates for the avid gardener group increased 6 percent year over year. The casual buyers group, now receiving a different type of catalog, exhibited similar increases.

In another example, a large industrial catalog initially was skeptical that statistical modeling would outperform its own intensive testing methods (it had tried it before, without positive results). It turned to a different vendor for a fresh perspective.

The new model that was created accurately predicted the best- and the worst-response segments with a nearly perfect ranking. In addition, the top-half segments outperformed the bottom-half segments significantly. Models subsequently have been developed for other product lines showing similar results in predictability.

A woodworking catalog used modeling to reduce the number of pieces mailed by 13 percent and increase net profits by more than 45 percent. Moreover, rather than use a single, previous mailing to predict a single, future one, the company combined the results of three previous mailings to predict a single, future mailing. This resulted in even higher response rates, and of course, profits. The model accurately had predicted a 47 percent response rate from the combination model’s top segment.

A home and garden catalog developed models to take better advantage of co-op databases and other prospecting efforts. It also developed upsell and cross-sell models for managing phone specials using analytics, rules and scripts for highly targeted product offers. Information pops up at just the right time as customer service reps (CSRs) interact with customers. In this way, CSRs are much better equipped to carry on seamless conversations armed with critical customer intelligence, with the same technology applicable to the company’s Web site.

Typical Objectives

Next, let’s look at typical objectives of direct marketers in applying statistical analyses and modeling to their operations and marketing efforts. Requirements for campaign forecasts include:

- allocation of all unknown demand, orders and units from each channel to the appropriate campaign (i.e., non-source-code items from phone and Web);
- forecasts for response percent, number of orders, number of units, total demand, average order dollars and sales per thousand catalogs by source code, customer segment, channel, and totals on a week-by-week and day-by-day level (factored for circulation — you wouldn’t mail all names in each cell);
- forecasts allocated by channel;
- pre-season forecasts of aggregate demand (and percent-complete) on a week-by-week basis, showing forecasts weekly and cumulative orders and dollars, with the ability to drill down to the book, key code, source code or product level;
- revised forecasts mid-season;
- response rate analysis accounting for responses by customers vs. prospects;
- forecasts and other metrics by customer segment, channel and totals compared to budget and prior year effort/results;
- forecasts that should include projected return and cancellation rates; and
- financial forecasts of net margin and profit and loss (projected vs. actual), which should incorporate data on fixed, promotional and fulfillment costs.

Other Considerations

Forecasts and comparative prior-period metrics should be broken down by response to product-type and catalog-page/ Web-page level. And for Web orders, set an offer/season different from the catalog.

To help improve your customer segmentation, define business rules for calculating customer RFM and product (RFMP) values. Those business rules should support aggregation of customer behavior by user-definable groupings/clusters for each dimension.

The RFMP business rules also should permit nested, multi-dimensional scoring, in which values assigned to each dimension can be placed in a user-defined algorithm to derive an aggregate score. The scoring’s results place a customer in a user-definable cell in a table, where the cells in turn can be clustered to derive a final RFMP score. The system should track automatically the evolution of each customer’s RFM score over time (i.e., migration based on customer behavior, not on user changes in the RFMP business rules).

Moreover, various segmentation and reporting criteria can be met by analytical modeling, such as order channel behavior (Web vs. catalog vs. retail); accounting for external demographic overlay data; and tracking the migration of customers from one segment to another over the full lifetime of the customer record.

Another tip: Select a group of customers by any user-defined characteristics, then save the selection characteristics for future reference and for re-use. Compare this group to any other group or statistically determined segment — or to your entire file — to determine distinguishing factors.

The most critical components here include customer level order tracking to analyze product purchasing behavior and product affinities. Customer level detail should include RFM designations and demographic data to allow for multiple views of the customer, including ZIP code/geographical area, seasonal purchase behavior, channel affinity, and original source. Also important: customer-level contact history through all channels, with the ability to cross these data with purchase history for campaign management applications. Generate any report at the customer-segment level (e.g., P&L by segment), and exclude specific customers, groups or segments from a report.

Solution Vendors

One of the best solution providers that can help with these tasks is ASA Corp. (www.asacorp.com). Its Customer Opportunity Advisor, Decision-Builder, ModelMax and ScorXPRESS products offer a full range of database marketing and analysis tools.

An equally catalog-oriented solution is RetailQuadware, now managed by Junction Solutions (www.junctionsolutions.com). It provides complex data mart and analytical solutions for catalogers such as Musician’s Friend, Johnson Smith and Levenger.

Others vendors to consider include DirectLogic Solutions (www.direct-logic.com) and Miglautsch Marketing (www.migmar.com). Several large service bureaus may be of interest, including: Harte-Hanks (www.harte-hanks.com), Experian (www.experian.com), and KnowledgeBase Marketing (www.knowledgebasemarketing.com). And depending on your requirements, you may be well-served by Abacus (www.abacus-us.com).

Ernie Schell is author of “The Guide to Catalog Management Software” and director of Marketing Systems Analysis, which helps catalog companies specify and select order processing software. Contact him at: ernie@schell.com.