BG/NBD Model

The model captures two distinct behavioral dimensions:

• Transaction Process: While active, a customer's purchases follow a Poisson process with rate λ
• Dropout Process: After each transaction, a customer has a probability p of becoming permanently inactive
• This separation allows the model to distinguish between a customer who is slow to purchase and one who has churned
• The 'buy-till-you-die' name reflects this exact mechanism: customers buy until they randomly 'die'

Individual differences are captured through mixing distributions:

• Transaction rate (λ) varies across customers following a Gamma distribution
• Dropout probability (p) varies across customers following a Beta distribution
• This hierarchical structure 'borrows strength' from the population, enabling robust predictions for customers with sparse histories
• The Beta-Geometric combination gives the model its name and handles the unobserved churn event

Unlike subscription models, BG/NBD operates without observed churn:

• Customer inactivity is probabilistically inferred, not directly observed
• The model calculates P(alive | purchase history) — the probability a customer is still active given their observed behavior
• This makes it ideal for retail, e-commerce, and CPG where customers don't formally cancel
• The recency of the last purchase is the strongest signal for inferring dropout

The model's mathematical elegance enables efficient computation:

• Fader, Hardie, and Lee (2005) derived a closed-form expression for the likelihood function
• This eliminates the need for computationally expensive Markov Chain Monte Carlo (MCMC) simulation
• Parameters (r, α, a, b) are estimated via Maximum Likelihood Estimation (MLE) using standard optimization
• The closed form enables rapid scoring of millions of customers in production environments

The primary output is E[X(T) | history] — expected transactions over a future horizon:

• Given a customer's frequency and recency, the model predicts expected purchases in the next T periods
• This conditional expectation integrates over both the transaction rate and the probability of still being alive
• Outputs feed directly into Discounted Cash Flow (DCF) calculations for CLV estimation
• Often paired with the Gamma-Gamma model to add monetary value predictions

The model has been rigorously tested against real-world datasets:

• Original validation used the CDNOW dataset, a canonical benchmark in CLV research
• BG/NBD consistently outperforms simpler heuristics like Pareto/NBD in holdout validation
• Performance is measured by comparing predicted vs. actual transactions in a future calibration period
• The model's parsimony — only four parameters — reduces overfitting risk while maintaining predictive accuracy

The standard companion sub-model to BG/NBD for estimating monetary value. It assumes a customer's average transaction value follows a Gamma distribution across the population, while individual transaction values are normally distributed. This two-stage Bayesian approach accounts for spend heterogeneity—the fact that some customers consistently spend more per order than others—independent of their purchase frequency. The model conditions on observed average spend and transaction count to predict future per-transaction value.

A closely related BTYD model developed by Schmittlein, Morrison, and Colombo that uses the Pareto distribution (of the second kind) for the dropout process instead of the Beta-Geometric used in BG/NBD. Key differences:

• Models time-to-churn as continuous rather than discrete per-transaction
• Requires more computationally intensive numerical integration for likelihood estimation
• Often yields nearly identical predictions to BG/NBD in practice
• BG/NBD is generally preferred for its computational efficiency and closed-form expressions

The statistical engine underlying the BG/NBD transaction process. It models purchase counts as a Poisson distribution where the rate parameter λ varies across customers according to a Gamma distribution. This mixture captures the empirical reality that:

• Some customers are inherently heavy buyers (high λ)
• Others are light buyers (low λ)
• The Gamma distribution flexibly accommodates this unobserved heterogeneity The resulting marginal distribution is a Negative Binomial, which naturally handles overdispersed count data.

The churn component of the BG/NBD model. After each transaction, a customer has a probability p of becoming permanently inactive. This dropout probability is modeled as:

• A Geometric process at the individual level (chance of churn after each purchase)
• A Beta distribution across the population to capture heterogeneity in churn propensity The Beta distribution is the conjugate prior for the Geometric, enabling closed-form Bayesian updating as transaction history accumulates. This is what makes BG/NBD computationally tractable.

A class of probabilistic models designed for non-contractual continuous settings where:

• Customer 'death' (churn) is unobserved—you never know when someone has truly left
• Transactions can occur at any time, not just at renewal periods
• The goal is to predict future purchasing while accounting for latent attrition The BTYD family includes:
• BG/NBD (Fader, Hardie, Lee, 2005)
• Pareto/NBD (Schmittlein et al., 1987)
• MBG/NBD (Batislam et al., 2007) — adds a 'revival' mechanism All share the core insight that purchase frequency and churn risk must be modeled jointly.

The BG/NBD model is inherently hierarchical Bayesian in its construction. It operates on two levels:

• Individual level: Each customer has latent parameters (λ for transaction rate, p for dropout probability)
• Population level: These parameters are drawn from prior distributions (Gamma for λ, Beta for p) This structure enables borrowing strength—customers with sparse histories are regularized toward the population mean, preventing overfitting. The Maximum Likelihood Estimation (MLE) procedure fits the population-level hyperparameters (r, α, a, b) from the entire cohort's transaction data.