Customer-Life-Time-Value forecasting

This kind of predictive model is actually a more advanced form of churn model.

Basically, in a churn model, we are computing the probability that a given customer moves from the “normal customer state” to the “churned customer state”. Let?s illustrate this with a small graph:


The “” means “Average Customer life-time value” of all the customers inside the segment_s: this value must be given by the business analyst or the marketing team.

The forecasting of the CLV(Customer-Lifetime-Value) of an individual “Ind_i” that is currently inside the “active state” segment is:

…or, in a simpler way:


…where is the probability of churn in the next 3 months for the individual ““.
This probability is computed using a predictive model (that is typically built with TIMi).

Here is another chart that explains the more general case:

The chart above contains 4 states (high-value-customers, medium-value-customers, low-value customers, churned customers). These states can be defined using business rules (that is the most “natural” way of doing it) or these states can be the results of an advanced segmentation created with Stardust. You must create a predictive model for each possible transition. This means that, in this simple case, there are already 8 predictive models to create (at least 5). In the chart above, not all possible transistions are illustrated. For example, for the , you must create a predictive model that estimate the probability that a customer inside the state “Medium-value-customer” switches to the state “Low-value-customer” in the next 3 months. As you can see, the creation of the whole system that is used to forecast the real CLV (3 months in advance) of a customer is a very time-consuming process because it involves creating many (many!) different models (one for each transition). Hopefully, with TIMi, creating a predictive model is quite fast. This is why TIMi is the optimal solution when building a system to predict the real CLV of a customer.

Once all the different predictive models are computed, there exists inside our analytical ETL tool (Anatella) a small box that directly predict the CLV of each customer. This box is quite simple: For example, the predicted CLV (in a 3 month window) of an individual “” currently inside the state “Low-value-customer” state is:

The predicted CLV (in a 6 month window ? i.e. two time-steps in advance) of an individual “” currently inside the state “Low-value-customer” state is:

Of course, the accuracy of the prediction of the CLV is only as good as the accuracy of the “individual predictive models”that are composing the CLV-prediction-system. Most of the time, these “individual predictive models” have already a very poor accuracy thus you should not expect a very accurate CLV prediction (especially when you are using any other datamining tool than TIMi to create the “individual predictive models“).

Datamining vendors are often searching to add new functionalities inside their tool instead of improving the functionalities already provided. Indeed, its very difficult to improve prediction accuracy (i.e. to obtain a higher lift on the TEST SET) and it is a lot easier to provide a new (barely working) functionality like this “CLV estimation system”.