In the world of customer centric targeted marketing, the biggest challenge that marketers face is that of balancing relevance with return on the investment (ROI). Relevance implies that the targeted message speaks to the targeted customer, both in terms of content as well as timing. For example, if a bank reaches out to a customer with a message about taking out a mortgage at an attractive interest rate, it would be relevant only if the customer needs the mortgage in the near future. An irrelevant message can turn off potential customers and can do more harm than good.
Predicting the Right Timing
This problem is a predictive problem where one not only needs to predict the likelihood that the customer would require a mortgage, but also get the timing right. Target the customer too early with a mortgage offer when she is not yet ready for the offer, and it’s a wasted contact. Target the customer too late, and she would have taken out with mortgage with another bank. Propensity models can only predict the likelihood of someone taking out a mortgage and not suitable for the scenario described here. Time to event models on the other hand predict time dependent propensities and ideal for this scenario.
Handling Giveaways with Monetary Value
The situation becomes more complicated when the targeting accompanies an actual giveaway of monetary value. Consider a retailer who provides relevant discount offers to its customers. Since they are relevant offers, they are ideal tool to increase customer engagement and loyalty. For example, if a customer loves a special kind of yogurt, making an offer on that product to the customer is an excellent targeted marketing strategy. Chances are though that the person would have bought special yogurt even without the offer. So the ROI on this particular offer would be 0 in such a case.
Addressing Incremental Sales
ROI is dependent on the incremental sales. Consider 2 customers, A and B, where A is likely to buy 1 unit of a product and B is likely to buy 2 units of the same product. If giving a discount of $1 induces customer A to buy 1 additional unit, and customer B to buy no additional unit, then giving that discount to customer A would yield better ROI than giving the offer to customer B, even though both end up buying 2 units of the product each.
Getting Around Impractical Experiments
To predict incrementally, traditionally offered and hold out groups are created and uplift models are then built on the data thus collected. Such experiments are costly as well as impractical, considering it would require tens of thousands of experiments to generate the data required to build uplift models for all the thousands of products that are sold. Our coupon effectiveness indices models don’t presuppose existence of experimental data and work directly with the historical business-as-usual data. From this data, we can always find 2 customers who are identical in all respect, but one was given an offer and another was not. Thus we can extract experiment-like data and build the uplift models on that data.
It is worth realizing that when giving out targeted discount offer, unlike in case of targeted messaging, ROI on the discount dollar becomes equally (and possibly more) important as the relevance of the discount offer. Our research has repeatedly shown that the customers most likely to buy a product are usually also relatively less likely to spend additional dollars on the offered product. Thus there is a trade-off between relevance and ROI. Once a marketer realizes this, it becomes easy to build solutions which look at both the aspects before making a discount offer to the customer. This combination ensures that the marketer would get the best bang for their discount buck!