Buy-Till-You-Die (BTYD) Models

BTYD models are specifically engineered for environments where customers do not explicitly terminate their relationship. Unlike subscription services with a clear churn event, these models infer latent attrition from observation windows.

• Mechanism: The model treats churn as an unobservable probabilistic event that occurs after any given transaction.
• Key Insight: A customer who hasn't purchased recently is not necessarily dead; they may simply be in a long inter-purchase interval.
• Application: Essential for retail, e-commerce, and hospitality sectors where customer silence is the only signal of potential defection.

The core mathematical innovation of BTYD models is the decomposition of customer behavior into two independent stochastic sub-processes that operate simultaneously.

• Transaction Process: While alive, the number of transactions a customer makes follows a Poisson distribution with a latent rate parameter (λ).
• Death Process: After each transaction, a customer may become permanently inactive with a probability governed by a Geometric distribution with parameter (p).
• Heterogeneity: Individual differences in λ and p are captured by mixing distributions (Gamma and Beta), creating a hierarchical Bayesian structure that borrows strength across the population.

The Pareto/NBD model, introduced by Schmittlein, Morrison, and Colombo in 1987, is the canonical BTYD framework that established the mathematical foundation for all subsequent variants.

• Assumptions: A customer's transaction rate follows a Poisson process; lifetime follows an Exponential distribution; heterogeneity in transaction and dropout rates follows Gamma distributions.
• Output: Derives closed-form expressions for expected future transactions over any time horizon, conditional on observed recency and frequency.
• Computational Evolution: Originally solved via complex Gaussian hypergeometric functions; modern implementations use Maximum Likelihood Estimation (MLE) for efficient parameter recovery.

The Beta-Geometric/Negative Binomial Distribution (BG/NBD) model, proposed by Fader, Hardie, and Lee in 2005, addresses the computational complexity of the Pareto/NBD while preserving predictive accuracy.

• Key Modification: The dropout process is shifted from occurring continuously to occurring immediately after a transaction, making the death process a Beta-Geometric mixture.
• Advantage: Produces analytically tractable expressions that can be computed with standard spreadsheet functions, dramatically reducing implementation barriers.
• Performance: Empirical studies show the BG/NBD matches or exceeds Pareto/NBD accuracy on most real-world datasets while being orders of magnitude faster to estimate.

BTYD models are typically coupled with a Gamma-Gamma spending model to produce a complete Customer Lifetime Value (CLV) estimate that accounts for both transaction count and average order value.

• Independence Assumption: The Gamma-Gamma model assumes monetary value is independent of transaction frequency, allowing separate estimation.
• Spend Heterogeneity: Individual differences in average transaction value are captured by a Gamma distribution, with a shape parameter that shrinks extreme values toward the population mean.
• DCF Extension: The combined output can be discounted using a continuous-time discount rate to compute the net present value of future customer cash flows.

The predictive power of BTYD models derives from their ability to compute conditional expectations — the expected number of future transactions given a specific customer's observed behavioral history.

• Sufficient Statistics: The model compresses a customer's entire transaction history into three summary statistics: Recency (time since last purchase), Frequency (total repeat transactions), and T (total observation time).
• Bayesian Updating: As new transactions are observed, the posterior distributions of individual-level parameters are updated, refining future predictions.
• Probability Alive: A derived metric that estimates the likelihood a customer is still active at the end of the observation period, often used as a churn risk score.

The foundational Buy-Till-You-Die model that predicts future transactions by modeling two stochastic processes simultaneously: a Poisson purchase process while alive, and a geometric dropout process that determines when a customer becomes permanently inactive. Heterogeneity across customers is captured using Gamma and Beta distributions for the transaction and dropout rates respectively.

• Inputs: Recency, Frequency, and Time since first purchase (T)
• Output: Expected number of future transactions in a given period
• Assumes transaction rate and dropout probability are independent

A monetary value model used in conjunction with BG/NBD to predict the average spend per transaction. It assumes that a customer's average transaction value follows a Gamma distribution across the population, and that individual transaction amounts are normally distributed around that mean.

• Does not assume correlation between purchase frequency and monetary value
• Requires at least one repeat purchase for stable estimation
• Commonly applied after predicting transaction count to compute full CLV

The original BTYD model that predates BG/NBD, modeling the transaction process as a Pareto distribution of the second kind. Unlike BG/NBD, it assumes that dropout can occur at any continuous point in time rather than only after a transaction.

• More computationally intensive than BG/NBD
• Uses the exponential-gamma mixture for inter-transaction times
• Often yields nearly identical predictions to BG/NBD in practice

A discrete-time retention model that predicts the probability a customer remains active in each future period. It assumes a Geometric distribution governs the period of churn, while heterogeneity in churn propensity across customers follows a Beta distribution.

• Ideal for contractual or subscription settings with defined renewal periods
• Models churn as a 'death' event rather than a latent inactive state
• Provides period-by-period survival probability estimates

A model validation technique that ranks customers by predicted CLV, divides them into ten equal-sized groups, and compares predicted values against actual realized revenue for each decile.

• Reveals whether the model correctly identifies top-value cohorts
• A well-calibrated model shows monotonic decreasing actual CLV across deciles
• Often visualized with Lorenz curves and Gini coefficients to quantify concentration accuracy

A statistical framework that estimates individual-level BTYD parameters by borrowing strength from the population distribution. Rather than fitting each customer independently, it assumes parameters are drawn from common hyper-distributions (Gamma, Beta).

• Enables robust predictions for customers with sparse transaction history
• Uses Markov Chain Monte Carlo (MCMC) or Maximum Likelihood Estimation for inference
• Prevents overfitting by applying Bayesian shrinkage to extreme values