How to measure customer lifetime value
Customer Lifetime Value (CLV) is a fundamental business metric, yet many organisations struggle to measure it accurately. This is often due to simplistic approaches rather than a lack of adequate data. In this post, I outline key principles for measuring and building trust in CLV estimates—an essential step toward informed, customer-centric decision-making.
“Customer centricity is a strategy to fundamentally align a company’s products and services with the wants and needs of its most valuable customers. That strategy has a specific aim: more profits for the long term.”
A key insight in customer behaviour analytics is that customer heterogeneity matters - some customers spend more, more often, for a longer period of time. Any model that attempts to accurately estimate CLV needs must account for these dynamics. The best approach is one that balances the need for accuracy with the challenges and costs of complexity.
CLV is critical for developing a customer centric business. That is, a business that is able to focus its efforts on finding and keeping their best customers and beat the competition.
What is CLV?
Customer lifetime value is defined as -
The present value of the future cash flows attributed to the customer during his/her entire relationship with the company. - Farris et al 2010
It is the value you can expect to generate from a newly acquired customer. It is a forward-looking metric that combines revenues, costs and retention, discounting future cashflows into present value. It is related to, though not the same as churn propensity modelling and a different approach is generally required. And it is more complicated than simply multiplying averages.
(What’s the story?)
We need a model that tells the story of how your customers interact with your business. It should capture the important dynamics, such as churn, payment frequency and customer heterogeneity, without getting stuck in unnecessary granularity.
For example, suppose a charity uses direct mail campaigns to acquire new donors. Once acquired, the donors pay a regular monthly fixed amount until they cancel.
This is a so-called ‘discrete time, contractual setting’, because payments are a fixed amount paid at regular time intervals and you get a clear signal when the ‘customer’ has canceled. This simplifies the story we need to tell with our model. In a ‘continuous time, non-contractual setting’ (e.g. eBay), we need a somewhat more elaborate story.
In this example, the model needs to capture the response to the campaign, donation values, expected donor ‘lifetime’, and any costs associated with managing their payments. Each of these components can be described using components of a mathematical model. For donor lifetime, for example, we propose using a ‘shifted beta geometric’ model. These are great at describing customer tenure and, crucially, at capturing differences in behaviours across the customer base.
How can we trust it?
Model trust is earned in two ways: validation and explainability.
Explainability means we can unpick the prediction and answer the questions - ‘why is CLV $x?’ and ‘why is CLV on z higher than y?’. Validation means we can demonstrate the accuracy and stability of the model. In the case of CLV estimation, we need to show that the model can accurately project the value of a new cohort of customers long into the future.
Plot 1 shows how CLV for the charity example is calculated.
Each green block represents the amount value we get at each month after acquisition for a typical donor.
In month 1, we get the full donation value (minus any costs associated with the transaction), which is about $10.
In month 2, we get a smaller amount due to churn risk estimated by the model. As the months progress, the probability the donor is still active fades (with the colour).
Over the 24 month period if we add up the bars we get a 2 year CLV (about $100 in this case).
(Note that in practice we usually apply discounting and an infinite time horzon when calculating CLV).
Plot 2 shows the estimated cumulative donation value per donor.
The green line is the prediction. It gives the projected cumulative value for a typical donor within a cohort, 2 years into the future. By month 24 it says we should expected to have just over $100.
The black lines are the actual cumulative average donation values for a selection of validation cohorts.
This is simulated data, but these are the kinds of plots you should be looking for (and asking from vendors) when assessing CLV model performance.
How do we keep the model relevant?
To get solid estimates for long term retention behaviour, we would ideally use data from customers acquired some time ago. A common question then is - given how much our company has changed over the years, how can we ensure our predictions are relevant to new customers?
The trick is to balance the information in the older data with what we know about the most recently acquired customers. To do this, we use Bayesian models (and ‘Gaussian processes’ in particular) to capture how retention behaviour evolves over time. The model will draw on the experience with older customer cohorts, but is pulled to-do-date with the first or second period retention rates of the newest cohorts of customers.
How can we make predictions for new products or in new markets where we have little or no data?
This is a common challenge. As with the example above, you want to draw on the broader context, while making your predictions relevant, without inferring too much from a small number of noisy data points. And again, as above, our approach is to use Bayesian methods.
Hierarchical Bayesian models allow you to balance accuracy and stability. As you gather more data for the new products or in the new markets, the model will learn how these differ from the existing ones.
How do we use it to make decisions?
CLV is essential for guiding a wide array of business decisions, from marketing, to product development and price optimisation. It can be used as in input to marketing mix modelling budget setting and channel optimisation planning or to place a monetary value on new product initiatives.
Charity example - how much to spend on acquisition?
The charity is offered a prospect list, scored for response propensity. They can choose the top N most likely to respond prospects to target with a direct mail campaign.
Plot 3 shows the declining response propensity across the list. We can use our CLV estimate to determine how far down the list to go.
Combining response propensities with cost and CLV estimates, we get plot 4. This shows the campaign net return at different campaign sizes.
It shows that we can target the top ~9000 prospects before the marginal return of an extra mailing falls below the expected lifetime value.
If the objective is to break-even, while maximising total response, we can mail up to around 47k prospects.
Our 9k mailing should generate a cohort of 467 new donors. We can use CLV to identify the point at which these will have raised enough donation value to pay back the cost of the campaign.
Plot 5 shows this is achieved after 4 months, when the cumulative donations from the cohort is expected to reach around $15k.
Customer lifetime value is an essential metric for any customer centric organisation. Overly simplistic approaches cannot estimate CLV accurately and will lead to poor decision making.
At DS Analytics, we use advanced Bayesian modelling techniques to quickly generate accurate, trusted, explainable CLV measures for our clients. Though the modelling techniques are challenging, the setup time and data requirements are not. We’ve found this to be one of the biggest wins for our clients.
Get in touch to find out if we can help you.