Marketing analytics

Towards Predictive Customer Lifetime Value

What’s Expected Customer Lifetime Value?

Expected Customer lifetime value (eCLV) is a term which describes the net profit resulting from the future relationship with a customer. eCLV is an interesting concept as it will show you how much you can spend on acquiring a new customer. A calculation of eCLV generally sums the expected profits in future time periods and estimated retention rate. Most models will also discount the cashflow. To make it a bit more concrete, you can find a simplified calculation of the eCLV below. In this example we expect $100 of net profit each year and a retention rate that drops 25 percentage points each year. We end up with an expected CLV of $250. $100 * 100% (year 1) + $100 * 75% (year 2) + $100 * 50% (year 3) + $100 * 25% (year 4) + $100 * 0% (year 5) = $250

Pitfalls in traditional customer lifetime value

While this number we end up with is actionable (i.e. we know how much we can spend on acquiring the customer) it is not necessarily correct. This traditional definition of eCLV has a number of pitfalls. The most notable ones are uncertainty and missing details. The first aspect – uncertainty – will always be present, i.e. the retention rate or the customer’s spending ends up differently then expected. Uncertainty also exponentially increases with the expected lifetime of a customer; we might be able to give relevant predictions for the coming year, but it becomes much more difficult to do so for the coming ten years. The second aspect – not enough detail – comes down to aggregation. Many of these eCLV models aggregate too much and thus do not take into account the specificity of certain customers. I.e. the eCLV could be wildly different for a 35 year old mother of two versus an 18 year old student.

Predictive customer lifetime value

As theory in this domain has evolved we have come to use the term predictive CLV instead of expected CLV. In predictive CLV we use machine learning techniques in combination with time series modeling to take into account all the details of a lead. So if the lead is a 40 year old male, likes to do sports, has two kids and works a job in retail, the predictive CLV model will take all this into account in order to calculate the value of this customer. By doing so, the pCLV will give a much more correct view of the value of a specific potential customer as the value will be based on all the characteristics of a person. So instead of asking yourself “how much can I spend on acquiring one customer?” you can now ask yourself “how much can I spend on acquiring a customer who is a male of 40 years, who likes to do sports, has two kids and works a job in retail?” and end up with a much more realistic answer. So why is it relevant to take all these aspects into account? Say for example you are in the sports retail industry, it seems logical that hours of sports per week are relevant in driving CLV. The graph above validates this train of thought, but also shows another interesting aspect. If we wouldn’t take into account gender our model would average out the specificities of each gender, ending with a much flatter line (the red one). This almost flat line seems to suggest that hours of sports per week is not important when it comes to driving CLV and removes our understanding of the specificities of male vs. female behaviour. While we talk about the relationship between CLV, gender and hours of sports here, you can imagine that in a real life case much more variables come into play. Once we reach more then 3 / 4 different characteristics and want to take into account the possible interplay between these, it becomes almost impossible to model this accurately using traditional CLV methods. Predicted Customer Lifetime Value (pCLV) can play an important role in making the eCLV model much more accurate. In pCLV, specific machine learning techniques are applied to understand the relationship between all these variables and build a model that can use all this information to most accurately estimate the pCLV. When training such a model it is possible to estimate the level of certainty we have about a given pCLV. This information is very useful in order to use the output in an actionable way. Next to making more informed decisions, these objective measures on accuracy can also be used to further optimize the model.

Linking acquisition, retention & churn to predictive CLV

Within the field of marketing analytics the state-of-the-art has evolved quite rapidly. While topics such as acquisition, churn and retention each have their own specific best-practices in terms of modeling, the current techniques allow for a strong integration of these different analyses. The value of linking up these models is that it allows one to make strongly informed decisions. As the pCLV model can be made very detailed it allows to give very detailed information on current and potential customers. For example, it will be able to show whether or not it is wise to invest in trying to acquire a specific customer. It will show the value of a current customer, allowing one to custom tailor the retention strategy. Furthermore, it will show the potential value lost if a customer churns, which will help in the decision on whether or not a recovery intervention should be executed.

Low hanging fruit

Out of experience, we can say that if up to now no data science methods were applied to optimize a company’s CLV calculation, there is almost always room to improve it using more state-of-the-art methods. It is difficult to estimate upfront what the cost of a predictive customer lifetime model will be as this is influenced by a lot of factors (e.g. available data, quality of the data, business domain, volume of customer set). However, generally is quite fast to do a first explorative analysis to understand the business process, have a look at the available data and make an initial feasibility study on whether or not a custom tailored pCLV model can translate into business value for a company.