Building a Customer Experience Simulation, Part II

Yesterday’s entry showed that a simple simulation model, predicting only whether customers make a purchase in each time period, could generate meaningful statistics about customer value and status. Such a model can be generated with a few lines of code in an agent-based modeling system. (I did just this using StarLogo TNG, available for free from http://education.mit.edu/starlogo-tng. Building something similar in a conventional programming language or process simulation tool would have been vastly more difficult. The current StarLogo TNG is a beta release that was unstable on my computer, but a final release is due before the end of the year and may be worth a second look.)

The only variable required in the model is the probability of each individual making a purchase. Where would this come from?

In the simplest possible model, the probability value would be a constant: say, a 50% chance of any customer making a purchase in any month. But that’s probably a bit too simple. In most businesses, recent buyers are more likely than others to purchase again. Since our simple model can generate Recency and Frequency statistics, it’s reasonable to incorporate these into the probability function. All you need is some historical data that shows how actual purchases relate to those values. The result would be a much more realistic simulation of customer behavior.

This model would be quite useful. Companies could seed it with the actual number of customers in each RF cell and get accurate estimates of future purchases by period. They could assess the value of promoting each RF cell, taking into account both the immediate response rate and later purchases by the “reactivated” customers. Changes in other assumptions would show the results of variations in customer behavior or in acquisition volumes. Additional statistics such as revenue, promotion expenses and profits could be derived from the same base calculations, enabling simple forms of customer value optimization.

All from modeling a single cell!

Of course, a real Customer Experience Matrix has many cells. The primary difference in a multi-cell model is that the probability formulas are more complicated. In particular, predictions for different activities will be interrelated. For example, the probability of a customer requesting a refund is likely to be based on recent purchases as well as previous refund behavior.

Yet a multi-cell simulation still shares the basic features of one-cell model: it works at the individual level and accumulates results by period. So a single-cell model is a good place to start when designing a Customer Experience simulation. And, as we’ve just seen, it offers considerable value of its own.