Flattening response rates, a public backlash against unwanted mail, and the threat of regulatory handcuffs due to privacy concerns are all clouding the future of direct marketing.

The solution, everyone agrees, is more sophisticated targeting and analysis.

Real difference

Knowing which customers will respond to a mailing, which ones will default on their payments, spend more (or less), defect to the competition, or turn into loyal users can make a real difference to direct marketing efficiency – and the bottom line.

A customer database can help to determine how customers will behave in future. But probing a database for answers can be time-consuming and arduous. Posing the right questions can be just as tough as getting the right answers.

Advanced statistical tools such as custom predictive modelling are usually required to find meaning in complex data relationships.

Predictive model

A predictive model seeks correlations between the behavior of customers and their characteristics. It identifies the factors that are most predictive and weights them according to their relative contribution.

The weightings are combined into an overall probability estimate or score. Individual scores are then computed for each customer. Customers are ranked from high to low, with the highest scored customers representing the most eligible target group.

Support objective

Predictive modelling can support just about any marketing objective. Some of the more common applications include:

– Optimizing response rates – By mailing to only the high probability responders, while suppressing those least likely to respond, the same results can often be achieved for less money.

– Maximizing customer spending – Customers with the greatest profit potential can be identified and encouraged to move up the spending pyramid.

– Minimizing attrition – By recognizing which customers are at risk to desert, pre-emptive action can be taken to save them before they sever the relationship.

Dormant customers

– Reviving dormant customers – By targetting those customers most likely to reactivate, while excluding customers who would have come back on their own anyway, significant savings can be achieved.

– Strengthening customer loyalty – Customers about to reduce spending can be identified in advance and given extra special attention before their spending drops by a single dollar.

Whatever the objective, predictive modelling can improve the odds of reaching the right customer at the right time, and is a more precise segmentation method than other techniques such as profiling.

Instead of dividing customers into groups or cells based on arbitrary performance measures (such as frequency, recency and monetary value of past purchases), modelling ranks customers individually by their projected future performance.

The outcome of a predictive model is a statistically derived decision rule that leads to more exact selection decisions.

A cutoff criterion is established – usually a target response rate, an allowable cost per order, or some other measure of success – and customers expected to surpass it are selected, while the balance of the universe is safely excluded.

Unlike profiling, predictive modelling makes no presumptions about which variables best explain customer behavior. Nor does it consider all explanatory variables to be equally important.

Using regression analysis, modelling assesses the respective contribution of each candidate variable; explores their interactive effect; finds the optimum combination; and then weights each according to its relative influence.

The cost of developing a predictive model is often recovered in a single mailing.

Within a typical mailing universe, a model will find a group of customers or prospects in the top 10% likely to respond at two to three times higher than average, and a group in the bottom 10% likely to respond at only about one-third the average.

Eliminate worst

Eliminating the worst-performing responders will yield major cost savings – without sacrificing sales targets. The savings can be pocketed – or used to fund other marketing initiatives.

The companies that are using predictive modelling span the business spectrum; the only factor common to them is their reliance on database-driven communications.

Following are some examples of successful modelling applications:

1) A continuity book publisher was experiencing a troublesome rate of returns on a free trial offer for a set of children’s encyclopedias.

While the up-front response to the offer of a free book was extremely high, too many of the trial subscribers ended up cancelling their orders, making the program unprofitable.

To salvage the program, the publisher needed to identify those customers most likely to respond and continue receiving the series.

Two model equations were developed: one to predict the probability of response, the other to predict the probability of return. The resulting model showed that only 30% of the target group was profitable to mail.

2) In an effort to revive dormant card accounts, a major gasoline retailer developed a communication program that offered cardholders high-value coupons on future purchases.

Over one-half

The program was successful in reactivating more than one-half of the inactive cardholders who were mailed the offer.

But research studies revealed that 40% of the target group might have resumed card spending on their own without any incentive.

If those cardholders could be identified and excluded from future mailings, major savings could be achieved.

The information known about each cardholder included summary level spending and payment data, as well as monthly transactional detail for the past two years.

The objective of the model was to establish correlations between all of these variables and the likelihood of a cardholder to resume card usage on their own.

Prime group

The model found a group of cardholders whose likelihood to reactivate spontaneously was twice the average and six times higher than the lowest group.

3) A major Canadian publisher wanted to improve the performance of a subscription promotion campaign for an upscale consumer magazine.

With a large stable of publications, this publisher could cross-sell the magazine to an extensive subscription base.

But the magazine catered to an affluent audience that could not be easily targetted using traditional list segmentation methods.

In an effort to lift response rates and reduce acquisition costs, a predictive model was developed to identify those active subscribers to other publications who would be most likely to respond.

A rich repository of data were available to analyze, including subscriber characteristics.

More than twice average

The model succeeded in finding a group of subscribers whose response probability was more than twice the average and who were seven times more likely to subscribe than the lowest responding group.

The best responding prospects could be acquired at one-third the average cost, whereas the worst responding prospects could only be acquired at a net loss.

As these examples demonstrate, custom models can lift response rates and improve profits.

With smarter targetting, mail volumes should recede – and consumer tolerance of direct mail will grow.

Stephen Shaw is vice-president of marketing and sales for Spectrum Decision Sciences, a Toronto firm specializing in custom analytical solutions for marketing.