Customer Lifetime Value: a business perspective

Customer Lifetime Value, also known as CLV or LTV, is more than a metric. It is a way of thinking about customer value across the company. Adam and Marek explored it in a Penguin Webinar, our educational webinar series. They described it as a tool that helps companies understand the value of their customer base and run the business with long-term sustainability in mind. CLV provides a framework for seeing how different customer segments contribute to growth. It can also help companies allocate budgets, plan product development, and optimize retention strategies.

CLV treats each customer as a long-term economic relationship rather than a one-off transaction. Instead of showing only how much someone has already spent, it estimates the value they may bring in the future. That forward-looking view helps CEOs, CFOs, and marketing directors make more confident decisions about segmentation, acquisition, and retention.

Why company leadership should care about CLV

CLV offers plenty of practical insight. It shows where investment is likely to pay off, which customers are most likely to be profitable, and where the strategy needs a rethink—perhaps because acquisition costs are rising unsustainably or new customers are proving less valuable. Many key decisions still rely on intuition. CLV grounds that intuition in data and gives leaders a clearer view of both current marketing performance and what lies ahead. It helps them ask the right questions:

How do customers from different sources differ?
Which segments deserve more attention and budget? Where should we be cautious or rethink the current strategy?
What is the value of our customer base?
What is the right price to pay for acquiring a new customer? Are we basing that decision on a one-off profit?
Are we financing growth with debt and falling into a discount trap?

Here CLV acts as a compass, helping you navigate varied data and turn it into concrete action. It is more than a number: it is a tool for managing growth strategically. It also bridges the gap between analytics and the executive team by giving data specialists and company leaders a shared language.

In the webinar, Adam revealed that his favorite use of CLV is spotting the discount trap. Companies often run discount campaigns to boost short-term revenue. CLV can show whether a customer who would have bought anyway simply ordered earlier and paid less because of the promotion. If CLV does not increase after the campaign, it clearly failed to have the intended effect.

Marek gave the example of a mobile app: app customers are usually more valuable than website customers. But what happens when we persuade website visitors to download the app? Does the download increase their value, or have they simply switched channels without spending any more overall? We do not know until we calculate CLV.

Adam then gave two less conventional examples. Imagine a retailer using CLV to tailor its offers: it might give its highest-value customers a personalized discount on products they buy often while showing everyone else the standard offers. CLV helps identify the customers worth that extra investment. The second example is inventory management. Some products may not sell much, but if our most valuable customers buy them, we make sure to keep them in stock. We do not want those customers shopping elsewhere and leaving us for good.

How CLV is calculated and why it is better to start simple

The webinar made one thing clear: CLV is not one specific number or a single calculation. It covers a range of approaches with different levels of complexity, data requirements, and interpretability. The right approach must match the company's maturity and the data it has available.

Before we start calculating…

Heuristics and ARPU

Some companies start with a calculation based on average order value and purchase frequency, or even just "the sum of revenue to date." Historical revenue describes what a customer has spent so far. Period-based ARPU (Average Revenue Per User) is average revenue per customer over a defined period. Neither is predictive CLV, which estimates future value over a stated horizon. These simple measures can still help with initial segmentation and show how much customer value differs across the company.

Segmentation (e.g. RFM)

Before any CLV calculation, segmentation is often used, for example RFM (Recency, Frequency, Monetary), which provides basic insight into customer behavior. Even without an advanced model, RFM can guide marketing, retention strategies, and customer service priorities.

Options for calculating CLV

Probabilistic models (BG/NBD, Pareto/NBD)

These models build on the principles of RFM and, thanks to their simplicity, have historically been among the most widely used in e-commerce. They model the probability that a customer will remain active and estimate how many more purchases they will make. They are often combined with a Gamma distribution to estimate order value. They work well for a PoC (proof of concept) and can produce a solid prediction even with limited data.

The model has its drawbacks, though. For example, it assumes that order frequency and order value are independent. That can be a serious problem: a higher frequency of lower-value orders, or vice versa, may prevent the model from capturing CLV declines caused by discount campaigns.

💡 Discounting: future profits can be discounted to their present value using an interest rate. In practice, however, this approach is used very little: it is too academic and not very practical.

Markov chains

Tipsport also experimented with Markov models, which calculate transitions between customer segments. This approach is not very widespread though, and they did not stick with it in practice.

Regression models and machine learning

For companies with more data, regression models can be useful, for example stochastic gradient boosting (LightGBM). They can predict future contribution margin directly, using dozens of input variables such as channel, product, demographics, and behavior. Here, contribution margin means revenue less the variable costs directly associated with serving the customer; it is a margin basis, not revenue or accounting profit. These models can be more accurate but harder to interpret. This approach is currently used by Tipsport.

AutoML

Cloud tools such as AutoML (e.g. on Google Cloud) let you test CLV prediction models quickly without a major development investment. The advantage is speed; the disadvantages are their black-box nature and potential data protection issues.

Hybrid approaches

Companies can combine probabilistic models with ML algorithms, for example predicting purchase frequency with BG/NBD and purchase value with regression. These approaches are the most complex and require advanced analytical skills.

Shorter prediction horizons

For shorter prediction horizons, working with pure CLV usually doesn't make practical sense. Most companies calculate customer value for the next year or two. In gaming, this period is extremely short: 7/30 days.

💡 CLV based on similar customers: CLV can also be evaluated based on customer similarity. This uses either a custom model, or it is part of ML models that work not only with a customer's order profile but also with their behavior.

This can introduce more randomness into the CLV model, so we recommend it only with very large customer datasets or when few dimensions are available. In gaming, for instance, the approach is simpler because the main differences in behavior come down to a handful of metrics: how long someone plays and what they do in the game. In e-commerce, however, customer behavior and CLV vary widely by the category or product purchased, which can leave similarity data rather sparse.

A clear comparison of the individual approaches to CLV:

Method	Advantages	Disadvantages	Statistical approach	Data needed	Assumptions
Heuristics (historical revenue or period ARPU)	Extremely simple, minimal data requirements, quick to calculate, easy to explain to management.	Ignores churn and future behavior; it is not predictive CLV and is sensitive to the observed period.	Descriptive calculation of historical revenue or average revenue per customer for a defined period.	Total revenue and number of orders per customer over a chosen period.	The past average order value and frequency stay constant; no margin or discount adjustment.
RFM segmentation	Quick insight into behavior, makes it possible to segment and target campaigns, easy to interpret.	Doesn't give a continuous CLV, depends on ad-hoc segment boundaries, ignores the trend over time.	Descriptive segmentation (quantile scores for Recency, Frequency, Monetary).	Date of last purchase, number of purchases, total spend per customer.	The segment score reliably represents future behavior; the boundaries are stable.
Probabilistic models (BG/NBD, Pareto/NBD + Gamma-Gamma)	Proven theory, closed-form analytical solution, works with a small amount of data, relatively interpretable.	Assumes frequency and value are independent, sensitive to marketing interventions, more complex implementation.	Bayesian hierarchical model: Beta-Geometric dropout, Poisson/Gamma heterogeneity, Gamma-Gamma for order value.	Transaction log (date and value of each order) for every customer.	Poisson purchase process, Beta distribution of churn probability, stationary behavior, frequency and value are independent.
Discounted cash flow (DCF)	Accounts for the time value of money, aligned with financial metrics NPV/IRR.	Sensitive to the discount rate estimate and to cash flow, doesn't solve prediction itself, rarely used in marketing practice.	Financial calculation of the net present value (NPV) of future cash flows.	Predicted cash flow by period, discount rate.	Constant discount rate, accurate cash flow estimates, risk captured in the rate.
Markov chains	Models multi-state behavior (active, passive, churn…), makes it possible to simulate the effect of interventions; the result is a CLV distribution.	Needs enough data for transitions, the Markov property may be violated, more complex to implement.	Discrete (semi-)Markov chain with rewards.	Sequence of customer states over time and the revenue associated with the states.	The future state depends only on the current one, transition probabilities are constant (or known over time).
Regression / ML models (GBM, Random Forest, NN)	High prediction accuracy, works with tens to hundreds of input variables, captures non-linearities and interactions.	Require a large amount of quality data, harder to interpret, risk of overfitting, higher computational cost.	Supervised machine learning regression models (e.g. Gradient Boosting, XGBoost, LightGBM).	Customer-level data mart: transactions, demographics, product and marketing attributes; target = future contribution margin over a stated horizon.	Training data represents the future, stationary distribution, sufficient feature engineering, and regularization.
AutoML platforms	Quick prototyping without a big investment, automatic model optimization, easy scaling in the cloud.	Black-box character, less control over the model, potential data protection and cost questions.	Ensemble of different ML algorithms managed by a cloud provider (e.g. Google Vertex AI, AWS SageMaker).	The same feature sets as for ML models, in a structured table (CSV, BigQuery, Snowflake…).	Quality and balanced data, the default pipeline is suitable for the domain, hyperparameters are tuned automatically.
Hybrid approaches (BG/NBD + ML)	Combines the interpretability of a probabilistic model with the accuracy of ML for order value, higher robustness.	Highest complexity, demanding on know-how and maintenance, longer development.	Composite: a probabilistic frequency model + an ML regressor for monetary value, possibly Monte Carlo simulation.	Transaction history as for BG/NBD + an extended feature set for spend modeling and simulation.	Validity of the assumptions of the individual sub-models, correct integration of outputs, stability of segments over time.

Start with segmentation

Segmentation is one of the key prerequisites for CLV; without it, there is little point in starting the calculation. The webinar made clear that companies should have their basic data preparation in hand, including RFM (Recency, Frequency, Monetary) segmentation, before getting into CLV modeling. RFM provides a simple but powerful framework for understanding customer behavior. Even without advanced models, it can guide marketing, retention strategies, and customer service priorities.

CLV is then often calculated at the segment level. A group of customers who bought recently, buy often, and spend more will differ from a group whose last purchase was a year ago; their estimated value will differ too. Segment-level CLV is usually simpler, more scalable, and often more accurate than trying to calculate it for each customer, especially when the data is sparse or noisy.

To begin with, we recommend focusing on larger customer clusters and assigning value to whole groups. In practice, that can mean splitting customers by acquisition source (e.g. organic vs. paid), type of first order (e.g. a sports category), or device (web vs. app). Companies can then compare which segments deliver higher CLV and adjust budget allocation or campaign targeting accordingly.

CLV and segmentation simply go hand in hand. The value of a customer acquired through the app may differ from that of a customer from a discount campaign, and it is segmentation that highlights this difference. If, for example, one customer segment responds much better to a retention program or has a longer life cycle, it makes sense to work with it differently than with less profitable groups. At Tipsport, for instance, they segmented by interest in a specific sport, because customers with different sports preferences may (but need not) have a different value.

At the same time, for some types of companies it is better to calculate CLV at the individual level. For example, "outlier" groups, meaning customers with extreme behavior or those who use only bonus promotions, can distort the averages of a whole segment. In that case it is worth creating a separate model or approach. The goal is always to balance model complexity with how useful the results are for business decisions.

Finally, there is value-based segmentation, which groups customers by their economic contribution. If a company calculates CLV at the individual level, it can use that value to build segments and set priorities across sales and marketing. This approach combines the power of predictive analytics with the clarity and simplicity of classic segmentation.

Marek and Adam in a heated discussion about the best approaches to CLV.

Where CLV helps—and where it doesn't

CLV is a useful tool for prediction and strategic management, but only when used in context and with a clear understanding of its limits. The webinar and the discussion that followed raised several important warnings about when CLV helps and when it can mislead.

First, CLV is an estimate, not 100% certain. It rests on historical behavior and assumptions about the future, and prediction accuracy drops significantly as the time horizon grows. A forecast covering 30 days to a year can be relatively reliable; a "lifetime" horizon of several years always carries uncertainty. Verifying a model over such a long period is practically impossible. Good predictions also require good, consistent data. If a company has sparse, low-quality, or inconsistent data, the outputs will be inaccurate. In that case, we recommend focusing on segmentation and heuristic approaches rather than relying on detailed models.

CLV also does not make sense for every type of business, including companies with very low purchase frequency or mostly one-time customers. At Tipsport, that can happen around elections or exceptional sporting events. There is simply nothing to predict. The same applies to businesses with low retention or customers who buy once a year: a lifetime-value model may add little. Segmentation or simple rules can be more useful in those cases. Even then, low average purchase counts can mask meaningful differences between acquisition channels or purchased categories.

Problems also arise when a company has not mastered the analytics basics, such as RFM segmentation, or lacks a clear use case for calculating CLV. The result may be an analytical output that nobody uses. Calculating CLV too early can have the same effect. Right after a customer's first purchase, models are unstable and their results tend to mislead. ML models can partly address this if we feed them behavioral and demographic metrics alongside the e-commerce data.

In fast-growing companies, where the acquisition mix changes quickly, models trained on historical data can fail, because new customers behave differently from those in the past.

The limitations lie not only in the type of business but also in the method used. A plain sum of revenue to date is historical revenue; it does not reflect the future. Period-based ARPU is an average for a defined period. A predictive CLV model must state its horizon and whether it predicts revenue or contribution margin; these are different quantities. Segment-level calculations can be misleading if a segment is too small or if a single "extreme" customer distorts its average. Probabilistic models such as BG/NBD or Pareto/NBD assume independence between frequency and order value, which often does not hold in practice. Because of this they also capture the effects of discounts poorly.

More complex models, in turn, are hard to interpret. It is difficult to explain why a model assigned a particular value to a particular customer, and this opacity can be a big obstacle in internal communication. AutoML compounds the problem: it produces results quickly, but often behaves like a black box. A similar complication arises when evaluating the effects of campaigns without A/B testing. Without a control group, you cannot tell whether CLV growth came from the campaign or from the underlying trend. And if we include costs in the model, for example for incentives or reactivation, we dramatically increase both the complexity and the risk of misinterpretation.

For these reasons, CLV is best treated as a useful framework rather than the one universal, definitive KPI. Its value grows with a company's data maturity, a clearly defined purpose, and realistic expectations about accuracy and usability.

What to take away

CLV is not just about the calculation. Above all, it is about thinking strategically. If you are wondering how to use it to set priorities, make investments, or build segments, we are happy to share what we have learned across different projects. Every business is different, but the path from data to meaningful decisions often follows familiar patterns.