Gamma-Gamma Model

The Gamma-Gamma model is a probabilistic model that estimates how much money a customer will spend per transaction — their expected average order value — as one half of a customer-lifetime-value calculation. It takes each customer's past purchase history and predicts a personalized future average transaction value, shrinking noisy customers with few orders toward the overall average and trusting heavy buyers' own history more. On its own it says nothing about how often someone will buy or whether they are still active; it answers only the monetary question, 'when this customer does buy, how much will the order be worth?' The model earns its name from its two gamma distributions: it assumes each customer's average spend is drawn from a gamma distribution, and that spend varies across the customer base according to another gamma.

The Gamma-Gamma model rests on a few assumptions worth stating plainly, because they decide when it is safe to use. It assumes a customer's order values scatter randomly around their own personal average, that this average stays roughly constant over the customer's life, and — crucially — that how much a customer spends per order is not correlated with how often they buy. That last assumption lets you model spend and frequency separately and multiply the results. If your heavy buyers also happen to place much larger orders, the independence assumption is strained and the estimates drift. Peter Fader, Bruce Hardie, and colleagues formalized the model for exactly this modular use: predict monetary value here, predict purchasing there, and combine them into a lifetime-value figure per customer.

The Gamma-Gamma model is almost always used with the BG/NBD model, and the two answer different halves of the same question. BG/NBD — the Beta-Geometric/Negative Binomial Distribution model — predicts behavior: how many purchases a customer will make in a future window, and how likely they still are to be 'alive' rather than quietly churned. Gamma-Gamma predicts money: given that a customer buys, how much each order is worth. Neither is a lifetime-value model by itself. You get customer lifetime value by combining them — expected number of future transactions from BG/NBD times expected value per transaction from Gamma-Gamma, discounted over the horizon you care about. Think of BG/NBD as the 'how often and how likely' engine and Gamma-Gamma as the 'how much' engine bolted onto it.

Keeping the two straight matters because they fail in different ways and are validated differently. BG/NBD leans on the pattern of purchase timing — recency and frequency — to judge whether a customer is still active, so it struggles in contractual settings where you already know who churned. Gamma-Gamma leans on the spread of order values and the independence of spend from frequency, so it struggles when big spenders also buy most often. A lifetime-value estimate is only as sound as both halves: a beautiful frequency forecast multiplied by a shaky spend estimate still gives a shaky number. When a CLV figure looks wrong, the discipline is to interrogate each model separately — is the frequency off, or the monetary value? — rather than blaming the combined output.

Use the Gamma-Gamma model where its assumptions roughly hold: a non-subscription, repeat-purchase business — most e-commerce and retail — where you can observe multiple orders per customer and where spend and frequency are not tightly linked. Before trusting it, check that independence assumption on your own data; if the correlation between order value and purchase frequency is meaningful, the monetary estimates will be biased and you should segment or model differently. Feed it customers with enough repeat history for the personalization to mean something, and treat one-time buyers with appropriate caution, since the model has little to learn from a single order. Then combine its per-customer average order value with a BG/NBD frequency forecast to build the lifetime-value figure you actually wanted.

The failures start with using the output as if it were the whole of CLV. The Gamma-Gamma model gives expected order value only; multiply it by frequency and survival, or you have half an answer. Applying it to subscription or contractual businesses, where the frequency logic does not fit, produces confident nonsense. Ignoring the spend-frequency independence assumption when your best customers plainly buy both more often and bigger will bias every estimate. And trusting per-customer figures for thin-history customers overreaches what the data supports. The discipline is to respect the model's narrow job — predict monetary value, nothing more — check its assumptions against your data, and combine it honestly with a frequency model rather than treating either half as the finished lifetime-value number.

The Gamma-Gamma model, named for its two gamma distributions, was formalized by Peter Fader, Bruce Hardie, and colleagues to estimate customer monetary value within lifetime-value modeling.

Etymology: source.

Search interest for this term over the last five years:

View interest-over-time on Google Trends →

Areas of marketing where gamma-gamma model is a core concern: