Life Time Value (LTV)

A customer is worth more than one purchase. The page starts with probabilistic and Bayesian lifetime-value models, then the libraries that implement them, and ends with the classic Pareto/NBD model explained conceptually.

The same notes are in Survival Analysis.

The modern starting point is probabilistic LTV using churn and neural nets, and Google's code for it is GitHub Google LTV, the google/lifetime_value repository. The older statistical line has its own library: lifetimes is lifetime value in Python. Between the two sits Bayesian Customer Lifetime Values Modeling using PyMC3.

Those libraries rest on the Pareto/NBD model, which the literature explains mostly in mathematics. Join Cross Validated: Understand Pareto and NBD models is the Cross Validated question asking for a simple, conceptual explanation of Pareto/NBD, from someone using the BTYD package to predict when a customer is expected to be back.

The two modeling approaches from the start of the page each have a write-up. Probabilistic LTV using churn and neural nets is https://towardsdatascience.com/the-paper-a-deep-probabilistic-model-for-customer-lifetime-value-prediction-eb5d61a83ecd, Maja Pavlovic's run through the neural network architecture and loss function of "A Deep Probabilistic Model for Customer Lifetime Value Prediction". Bayesian Customer Lifetime Values Modeling using PyMC3 is https://towardsdatascience.com/bayesian-customer-lifetime-values-modeling-using-pymc3-d770676f5c06, Meraldo Antonio's implementation of BG-NBD, a probabilistic hierarchical model, in PyMC3 to analyze customer purchase behavior.