Pareto/NBD Model

Because the Pareto/NBD lets customers drop out at any moment in continuous time, its marginal likelihood, obtained by integrating over both the gamma-distributed buying rate and the gamma-distributed dropout rate, contains a Gaussian hypergeometric function. Evaluating this special function accurately and stably across the parameter space is numerically delicate, and naive implementations can fail to converge or converge to poor solutions. The BG/NBD sidesteps much of this by assuming dropout can occur only immediately after a purchase, yielding simpler closed-form expressions that can even be computed in a spreadsheet. This practical estimation difficulty, rather than any deficiency in the model's behavioral story, is the main reason analysts often reach for the BG/NBD first.

P(alive) is the model's estimate of the probability that a given customer is still an active buyer at the end of the observation window, given their recency, frequency and tenure. It balances two signals: high frequency suggests an engaged customer, but stale recency, under the continuous-time death process, is strong evidence that the customer has churned. So a customer who used to buy often but has been silent for a long time relative to their pace will receive a low P(alive), while a customer whose latest purchase is recent relative to their history will receive a high one. Managers use P(alive) to rank customers by liveness, detect valuable customers slipping away, and prioritize retention spending, but it should be read as a model-based probability, not a certainty.

No. Like other buy-till-you-die models, the Pareto/NBD is exclusively about transaction timing and counts: it predicts how many future purchases a customer will make and whether they are still active, not the value of those purchases. To obtain monetary customer lifetime value you combine the Pareto/NBD's expected-transactions output with a separate spend model, most commonly the Gamma-Gamma model, which estimates each customer's expected average transaction value. Multiplying expected future transactions by expected spend and applying a discount rate yields residual lifetime value. Keeping 'how often' and 'how much' as separate, modular models is a deliberate and standard design choice in customer-base analysis.