Beta-Geometric/Negative Binomial Distribution (BG/NBD) Model

The beta-geometric/negative binomial distribution (BG/NBD) model is a probability model that predicts how often a customer will buy in the future, and whether that customer is still active at all, using only their past purchase history. It was introduced by Peter Fader, Bruce Hardie, and Ka Lok Lee in their 2005 Marketing Science paper, "'Counting Your Customers' the Easy Way," as a simpler alternative to the earlier Pareto/NBD model. It belongs to a family of buy-till-you-die models. The idea behind that vivid name is that each customer keeps purchasing at their own rate until, at some unobserved point, they quietly become inactive — they never tell you they have left. The model captures both behaviors at once: how frequently a customer buys while active, and how likely they are to have already dropped out.

BG/NBD is built for the non-contractual setting — the world of online stores, retailers, and many consumer apps where customers can lapse silently, unlike a subscription where churn is a visible cancellation. It assumes each customer has their own buying rate and their own dropout tendency, drawn from population-wide distributions (the negative binomial distribution for purchasing, the beta-geometric for dropout, which is where the name comes from). From a simple summary of each customer's history — how many times they bought, how recently, and over how long — the model estimates the expected number of future transactions and the probability the customer is still alive. That makes it a core engine for customer lifetime value, since lifetime value depends on how much a customer will buy before they go.

The everyday alternative to BG/NBD is RFM — scoring customers by recency, frequency, and monetary value — or a flat rule like "inactive after ninety days." Those are easy but crude: they bucket customers without a real model of buying and dropout, so they cannot give a calibrated probability that a specific customer is still active or a defensible forecast of future purchases. BG/NBD turns the same recency and frequency inputs into actual probabilities and expected counts. It is also the friendlier cousin of the Pareto/NBD model, which it was explicitly designed to approximate with far less computational pain — Fader and colleagues showed its parameters can be estimated even in a spreadsheet, which is a large part of why it spread.

A key honest caveat: BG/NBD models how many times a customer will buy, not how much they will spend per purchase. To get to dollar lifetime value you pair it with a separate spend model — commonly the Gamma-Gamma model — which estimates average transaction value. The two together produce a monetary CLV estimate. The model also rests on assumptions: customers act independently, buying and dropout follow the assumed distributions, and the future resembles the calibration period. When a business is genuinely contractual, with visible start and end dates, a survival or churn model fits the data better than a buy-till-you-die model designed for silent attrition. Match the model to whether your customers can leave without telling you.

Using the BG/NBD model well means feeding it clean per-customer history — number of repeat transactions, time of the most recent one, and the length of the observation window — then validating the fit before trusting it. Hold out a later period and check that the predicted number of transactions tracks what actually happened across customer groups; if it does not, the assumptions may not suit your data. Use the model's two outputs deliberately: the probability a customer is still active is a powerful churn-risk and reactivation signal, and the expected future transactions feed lifetime value and budgeting for acquisition and retention. Pair it with a spend model when you need dollars, not just counts.

Watch the traps. BG/NBD assumes silent, non-contractual dropout, so do not apply it to a subscription business where churn is observable — a survival model fits that far better. Do not read its transaction forecast as a revenue forecast without a spend model attached. Refit periodically, because customer behavior and the business itself drift, and a model calibrated on an old period can mislead. And resist over-trusting any single customer's prediction: these models are most reliable in aggregate, across the base, where the population-level patterns hold, and noisier for one individual. Used with those cautions, BG/NBD turns raw purchase logs into calibrated, decision-ready estimates of who will buy again and who has likely gone.

The beta-geometric/negative binomial distribution (BG/NBD) model (Fader, Hardie & Lee, 2005) predicts future purchasing and silent churn of non-contractual customers and is a workhorse for customer lifetime value.

Etymology: source.

Search interest for this term over the last five years:

View interest-over-time on Google Trends →

Areas of marketing where beta-geometric/negative binomial distribution (bg/nbd) model is a core concern: