BG/NBD Model
Also known as: Beta-Geometric/NBD Model, BG/NBD, Buy-Till-You-Die Model, Fader-Hardie-Lee Model
The BG/NBD (Beta-Geometric/Negative Binomial Distribution) model is a probabilistic buy-till-you-die model that predicts how many times a customer will transact in the future and whether that customer is still active, using only their past purchase recency and frequency. Introduced by Peter Fader, Bruce Hardie and Ka Lok Lee in their 2005 Marketing Science paper "Counting Your Customers the Easy Way," it was designed as a far simpler alternative to the Pareto/NBD model of Schmittlein, Morrison and Colombo while delivering comparable forecasts. The model couples a Poisson purchasing process, whose rate varies across customers by a gamma distribution, with a geometric dropout process governed by a beta-distributed dropout probability. The key behavioral story is that customers buy at a steady individual rate while alive and become permanently inactive with some probability immediately after any purchase. Because the latent attrition is unobserved, the model infers each customer's probability of still being alive from how recently and how often they bought. Its estimation requires only the (x, t_x, T) summary per customer and can even be fit in a spreadsheet, which made customer-base analysis practical for ordinary analysts.
Key highlights
- Requires only recency, frequency and tenure (x, t_x, T) per customer, so it runs on minimal transaction summaries and even fits in a spreadsheet.
- Has fully closed-form expressions for the likelihood, expected future transactions and P(alive), making estimation fast and reproducible.
- Delivers forecasts that are essentially as accurate as the more complex Pareto/NBD while being far simpler to implement and explain.
- Produces an intuitive, per-customer probability of being alive that directly supports targeting, retention and lifetime-value analysis.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use the BG/NBD model in non-contractual, continuous-purchasing settings where customers can lapse silently and you only observe a transaction log, such as e-commerce, retail, donations or many app contexts. It is appropriate when you can summarize each customer by recency, frequency and tenure and when purchasing is reasonably well described by a steady-while-alive Poisson process with one-shot dropout after a purchase. The model shines when you need to forecast future transactions, estimate P(alive), or rank customers for targeting and CLV, and when you want something materially easier to implement than the Pareto/NBD. It is less suitable for contractual settings (subscriptions, where churn is observed and survival models are better), for highly seasonal or trend-driven demand that violates the stationary-Poisson assumption, or when purchases are strongly time-clustered. Because BG/NBD models only the timing and count of transactions, it must be paired with a spend model such as Gamma-Gamma to produce monetary lifetime value.
Strengths & limitations
- Requires only recency, frequency and tenure (x, t_x, T) per customer, so it runs on minimal transaction summaries and even fits in a spreadsheet.
- Has fully closed-form expressions for the likelihood, expected future transactions and P(alive), making estimation fast and reproducible.
- Delivers forecasts that are essentially as accurate as the more complex Pareto/NBD while being far simpler to implement and explain.
- Produces an intuitive, per-customer probability of being alive that directly supports targeting, retention and lifetime-value analysis.
- Assumes dropout can occur only immediately after a purchase, so a customer who never makes a repeat purchase is treated as always alive, which can distort some forecasts.
- Relies on a stationary Poisson purchasing process and ignores seasonality, trends, marketing covariates and time-varying behavior.
- Models only transaction timing and counts, so it cannot by itself say anything about how much customers spend.
- Population-level heterogeneity is forced into gamma and beta forms, which may misfit bases with multimodal or otherwise non-standard behavior.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How does the BG/NBD model differ from the Pareto/NBD model?
Both are buy-till-you-die models that combine an NBD purchasing process with a latent dropout process, and both produce expected future transactions and P(alive). The key difference is the dropout story. In the Pareto/NBD, a customer can drop out at any moment in continuous time, with dropout times governed by a gamma-mixed exponential (Pareto) process. In the BG/NBD, dropout can only happen immediately after a purchase, with the dropout probability beta-distributed across customers. This discrete-time assumption makes BG/NBD far easier to estimate, with closed-form expressions that even fit in a spreadsheet, while delivering forecasts that are typically very close to the Pareto/NBD in accuracy. One quirk is that BG/NBD treats a customer who has made no repeat purchase as still alive.
What data do I need to fit a BG/NBD model?
Remarkably little. For each customer you need three numbers summarizing the calibration period: frequency x (the number of repeat transactions, that is, total transactions minus the first one), recency t_x (the time of the last transaction measured from the customer's first purchase), and T (the total length of time the customer has been observed, from first purchase to the end of the calibration window). You do not need the full transaction-by-transaction detail to estimate the four population parameters, and you do not need any demographic or marketing data. This minimalism is a major reason the model is so widely used; almost any business with a timestamped order history can construct the required (x, t_x, T) summary.
Can the BG/NBD model tell me a customer's lifetime value?
Not on its own. The BG/NBD model is purely about transaction timing and counts: it predicts how many times a customer will buy and whether they are still active, but it says nothing about how much money each purchase generates. To get monetary customer lifetime value you combine the BG/NBD's expected-transactions output with a separate model of spend per transaction, most commonly the Gamma-Gamma model of Fader and Hardie, which estimates each customer's expected average order value. Multiplying expected future transactions by expected spend and applying an appropriate discount rate yields a residual lifetime value. This modular separation of 'how often' from 'how much' is a hallmark of the modern customer-base-analysis toolkit.
Sources
- 1.Fader, P. S., Hardie, B. G. S., & Lee, K. L. (2005). "Counting Your Customers" the Easy Way: An Alternative to the Pareto/NBD Model. Marketing Science, 24(2), 275-284.
- 2.Schmittlein, D. C., Morrison, D. G., & Colombo, R. (1987). Counting Your Customers: Who Are They and What Will They Do Next? Management Science, 33(1), 1-24.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 23). BG/NBD Model. ScholarGate. https://scholargate.app/marketing/bg-nbd-model