How the Pareto/NBD and BG/NBD Models Predict Customer Lifetime Value in Non-Contractual Settings

For data scientists forecasting customer behavior, the choice of probability model forces a structural trade-off regarding human attrition. Advocates of the foundational Pareto/NBD model argue that customer churn is a continuous risk—a shopper might lose interest at any moment, meaning their probability of remaining active decays steadily over time regardless of their actions. Conversely, proponents of the Beta-Geometric/Negative Binomial Distribution (BG/NBD) model argue that churn is a discrete event tied exclusively to the purchase itself; a customer only decides to abandon a brand at the exact moment they complete a transaction. Neither side disputes the ultimate goal of predicting Customer Lifetime Value in non-contractual settings, but their differing assumptions about how and when a customer effectively "dies" yield drastically different financial forecasts.[1][2][3]

The necessity for these complex probabilistic frameworks stems from the fundamental ambiguity of non-contractual business models. In a contractual setting, such as a telecommunications subscription or a gym membership, a company knows exactly when a customer churns because the relationship is formally severed. In non-contractual environments like e-commerce, retail, or charitable donations, attrition is silent. A customer simply stops buying. Without a definitive cancellation event, analysts cannot easily distinguish between a loyal patron who is merely between purchases and a former customer who has permanently defected to a competitor.[6]

To solve this, statisticians developed the "Buy Till You Die" (BTYD) class of models. As defined by Wikipedia's documentation of the statistical class, these models are designed to "capture the behavioral characteristics of non-contractual customers, or when the company is not able to directly observe when a customer stops being a customer of a brand." By treating observed purchase history as the outcome of a random process governed by latent characteristics, BTYD models estimate the expected number of future transactions and the probability that a customer is still active.[6]

The foundational architecture for this approach was established in 1987 by David Schmittlein, Donald Morrison, and Richard Colombo with the introduction of the Pareto/NBD model. This framework relies on two distinct probabilistic processes running in parallel. The first governs the buying behavior: while a customer is "alive," their purchases follow a Poisson process with a constant transaction rate, denoted by the Greek letter lambda. This means that transactions occur independently of one another at a steady average pace.[2][6]

The second process governs the churn behavior. The Pareto/NBD model assumes that a customer's lifetime is unobserved and continuous, following an exponential distribution with a dropout rate denoted by mu. Because every customer is different, the model does not assume a single lambda or mu for the entire customer base. Instead, it assumes that these individual transaction and dropout rates are drawn from two independent Gamma distributions across the population, capturing the natural heterogeneity of a diverse market.[2]

Despite its theoretical elegance, the Pareto/NBD model presents significant computational challenges. Evaluating the likelihood function requires calculating complex hypergeometric functions, which historically made parameter estimation slow and numerically unstable, particularly for large datasets. While modern computing has largely mitigated these issues, the mathematical friction prompted researchers to seek a more tractable alternative that could deliver similar predictive power without the computational overhead.[1][5]

In 2005, Peter Fader, Bruce Hardie, and Ka Lok Lee published a landmark paper in the journal Marketing Science titled "Counting Your Customers the Easy Way." They introduced the BG/NBD model as a computationally lighter alternative. The innovation of the BG/NBD model lies in its slight alteration of the behavioral "story" associated with customer attrition. Rather than assuming that churn can happen at any continuous moment, the BG/NBD model assumes that a customer can only churn immediately after a transaction.[1][3]

Under the BG/NBD framework, the purchase process remains identical to the Pareto/NBD model—a Poisson process with a Gamma-distributed transaction rate. However, the dropout process is modeled as a Beta-Geometric distribution. After every purchase, the customer flips a weighted coin to decide whether they will ever return. The probability of dropping out is constant for that individual but varies across the population according to a Beta distribution. This discrete approach to churn eliminates the need for complex integrals, yielding closed-form likelihood expressions that are vastly easier to optimize.[3][5]

The practical stakes of choosing between these two models become apparent when they are applied to real-world data. Because the BG/NBD model forces churn to be evaluated only at the moment of purchase, it can aggressively penalize the survival probability of customers who have experienced long hiatuses since their last transaction. The Pareto/NBD model, which allows for continuous survival, often yields a more optimistic assessment of a customer's residual value, particularly in datasets where transaction counts are heavy-tailed.[2][3]

A comparative analysis using a 32-customer sample dataset from MetricGate illustrates this divergence starkly. When the Pareto/NBD model was fitted to this cohort using maximum likelihood estimation, it yielded a mean transaction rate of 0.19 per period. Over a 12-period forecasting horizon, the model projected a mean expected residual value of 22.78 currency units per customer, with an overwhelmingly high mean probability of being alive (P(alive)) at 0.9997. The documentation notes that "Pareto/NBD remains the gold standard for non-contractual customer base analysis" due to this flexibility.[2]

Fitting the BG/NBD model to the exact same 32 customers produced a radically different forecast. The model estimated a higher mean transaction rate of 0.33, but paired it with a mean dropout probability of 0.16 after each purchase. Consequently, over the same 12-period horizon, the BG/NBD model projected a mean P(alive) of just 0.0062. This pessimistic survival outlook collapsed the financial forecast, yielding a mean Customer Lifetime Value of only 0.44 currency units—a reduction of more than 98 percent compared to the Pareto/NBD output.[3]

This massive discrepancy highlights the sensitivity of probabilistic models to their underlying assumptions. For a marketing department relying on these figures to set Customer Acquisition Cost (CAC) limits, the choice of model dictates the budget. If the Pareto/NBD forecast of 22.78 is accurate, the company can afford to spend aggressively to acquire similar cohorts. If the BG/NBD forecast of 0.44 is closer to reality, that same acquisition spend would result in catastrophic losses.[7]

It is important to note that neither model natively predicts the monetary value of future transactions; they only predict the volume of future purchases and the likelihood of retention. To complete the Customer Lifetime Value equation, analysts typically pair these frequency models with the Gamma-Gamma model. The Gamma-Gamma extension assumes that the monetary value of a customer's given transaction varies randomly around their mean transaction value, which itself varies across the population.[4][6]

By multiplying the expected number of future transactions (derived from either Pareto/NBD or BG/NBD) by the expected average order value (derived from the Gamma-Gamma model), businesses arrive at the Discounted Expected Residual Transactions (DERT). This final figure represents the present value of all future cash flows associated with the customer, providing the ultimate metric for ranking customers by expected residual value for retention or upsell budgets.[2][3]

The evolution of Customer Lifetime Value modeling continues as researchers seek to bridge the gap between these probabilistic baselines and modern machine learning techniques. While the foundational models rely strictly on Recency, Frequency, and Monetary (RFM) history, contemporary extensions are beginning to incorporate time-invariant covariates—such as customer demographics or acquisition channels—to explain the heterogeneity among buyers. Until those hybrid models become standard, data scientists must continue to weigh the mathematical flexibility of the Pareto/NBD against the computational efficiency of the BG/NBD, knowing that the timing of a theoretical "death" determines the value of the living.[4][7]