When targeting is used to stimulate demand, the main problems that arise are taking credit for purchases that would have happened anyway, and giving away discounts and special offers needlessly. This is not to say that discounts should never be given to people who would have purchased without them; but if the goal of an incentive programme is to drive up the level of business, it is important to measure the programme on the basis of its true incremental impact, and to target the activity so as to maximize the increase in business generated.
Everyone is familiar with the idea of using "response modelling" to identify the characteristics of customers who "responded" to previous direct marketing activity: this is the main-stream state of the art. Because it is also recognized that not all of these "responses" (purchases) are incremental, it is also accepted best-practice to use control groups to assess the net impact of the activity. However, there is a disconnect when it comes to targeting the next campaign.
Perhaps the best known way of assessing the impact of targeting is to use a gains chart or gains table. These show, cumulatively, for each decile, the proportion of the sales generated when targeting the whole population that can be generated by targeting only the "best" x%. This leads to all the usual 80-20-type statements, such as "We managed to achieve 82% of the sales while targeting only 32% of the population".
The danger with this approach is that in most cases, that many of the sales achieved would have been made even without the intervention. Often the easiest way for a model to generate a good-looking gains chart is to rank highly people who have a high propensity to purchase independent of any marketing activity.
Of course, control groups are used precisely to avoid such double-counting or false credit taking. Given a properly randomized and appropriately structured control group, we can measure with a high degree of accuracy the value of extra sales a marketing programme is responsible for generating.
But while control groups can allow us to quantify the net impact and true profitability of a marketing campaign, they do not, in themselves, tell us how to optimize future targeting for maximum profitability. This is because standard "response" modelling doesn't, in fact, model true "response" at all: it can't because the controls are usually discarded completely during modelling. "Response" models are really purchase models, and they are completely unable to distinguish an incremental purchase caused (or made more likely) by marketing activity from one that would have would have happened anyway.
A new class of modelling approaches address this problem. They go under various names, including uplift models, differential response models, incremental impact models, and net models. They have been proved, in many cases — through in-market trials, longitudinal validations and cross-validations — to out-perform traditional response models by large margins. This is particularly true in applications in which high value products are being sold with low overall purchase rates and even lower overall incremental impacts from campaigns. Here, both traditional response models and the simplest approaches to incremental impact modelling fail, but careful use of modern approaches can yield powerful predictions that provably allow much more accurate targeting of those customers whose behaviour can be positively affected by a particular intervention. They can also be used to choose from among a number of possible interventions, optimizing the approach for each customer.
Staff at Stochastic Solutions have world-leading experience in these approaches, and have worked with companies at all stages of development and sophistication of approach. Stochastic Solutions can help evaluate your company's current approaches to customer targeting, and assess the potential impact that moving to more sophisticated approaches might bring. We can help you assess the validity and effectiveness of your current approaches to control group design, or if you don't currently use systematic controls, can help you to adopt a methodology that will allow you to evolve first towards mainstream best-practice, and then, if appropriate, on to the most effective, modern, scientific approaches to customer targeting.
