THE MODEL LINEAGE

Each generation of models widened the reality we can see.

From the 1959 NBD to today, the main line of consumer-behavior statistical models is only four steps. Most CRM implementations are still on step one.

Press → to advance along the lineage · A autoplay · F fullscreen

THE QUESTION

Hold your CRM system up against this lineage.
Which column is it actually in?

If the answer is 1959 to 1982, your customer insight is stuck at "sees Fate, can't see Fortune."
That is what this is really about.

CONSUMER-BEHAVIOR STATISTICAL MODELS · 1959-2026

The Model Lineage · what each generation can see

Timeline on top, capability matrix below. The same set of consumer phenomena, and different generations of models see different things.

HOW TO READ

Top half: the model timeline. Six nodes on the main line. Sharp branches off in 2010, a parallel track rather than the next step on the main line.

Bottom half: the capability matrix. Each row is a consumer phenomenon, each column a generation of models. The mark in a cell shows whether that generation can see that phenomenon.

● models it stably ｜ ◯ partially ｜ — can't see it

Press → to advance along the lineage.

STAGE 1 · 1959 - 1982

NBD and the River of Time: the start.

Ehrenberg's 1959 Negative Binomial Distribution is the mathematical start of all of this: a Gamma distribution for the purchase-rate differences between consumers, a Poisson for the randomness within one person.

Greene in 1982 wrapped that math in the "River of Time" metaphor, to make it usable for marketers.

This generation's achievement is stably seeing between-person heterogeneity, the differences of Fate.

●Between-person heterogeneity (different people, different λ)
—Change within one person (staying, exiting, accelerating, slowing, switching category)

EHRENBERG 1959 · CHATFIELD-EHRENBERG-GOODHARDT 1966 · GREENE 1982

STAGE 2 · 1987 - 2005

The Fader lineage: death enters the frame.

Schmittlein, Morrison, and Colombo's 1987 Pareto/NBD added a phenomenon Greene never modeled: consumers churn permanently. The model can estimate "the probability this customer is still alive."

Fader, Hardie, and Lee's 2005 BG/NBD simplified the estimation. Both are widely deployed commercially; Zodiac, which Fader co-founded, was acquired by Nike in 2018.

But the "while alive" λ is still assumed constant. Greene's person has not woken up.

●Between-person heterogeneity + permanent churn
—λ change within one living person (life-stage shifts stay invisible)

SCHMITTLEIN/MORRISON/COLOMBO 1987 · FADER/HARDIE/LEE 2005

STAGE 3 · 2008

Netzer: state switching.

Netzer, Lattin, and Srinivasan's 2008 HMM lets a consumer migrate among several latent states. Each state has its own λ. The transition probabilities are estimated from the sequence.

This is the first frame that seriously answers "change within one person." In his 1982 apologia Greene admitted his model could not do this. HMM does.

But the state variable is low-dimensional (usually 2 to 4 latent states), and fine-grained change across category, channel, and SKU still needs more complex models.

●Between-person heterogeneity + permanent churn + state switching
—Multi-dimensional state vector (category, channel, basket moving together)

NETZER/LATTIN/SRINIVASAN, MARKETING SCIENCE 2008

STAGE 4 · 2010 · PARALLEL TRACK

Sharp: the counter-proposition.

Byron Sharp (How Brands Grow, 2010) is another track rather than the next step on the main line.

His observation rests on decades of Ehrenberg-Bass mass-market data: in FMCG and everyday goods, consumer λ clusters in a low band, so between-person heterogeneity is actually not that large. Marketing resources should go to penetration and mental availability rather than individual differentiation.

Both tracks are right, it depends on the category. Luxury, premium, subscription, and life-stage categories fit the main line; mass everyday FMCG fits Sharp.

●Mass-market penetration and frequency dynamics (brand level)
—Individual customer state evolution (outside this frame's scope)

SHARP, HOW BRANDS GROW · EHRENBERG-BASS INSTITUTE · 2010, 2018

STAGE 5 · 2018 +

Modern sequence models: multi-dimensional state.

After 2018, RNNs and Transformers made the sequence structure itself the modeling object. The state vector can be multi-dimensional (frequency × category × channel × basket × inter-purchase time).

Modern sequence models stably capture category switching, life-stage shifts, and cross-category upgrades, the phenomena a single-variable HMM cannot see.

The cost of this generation is a sharp drop in explainability. HMM can still tell you "which state this person is in now"; a deep sequence model often gives you only an embedding vector.

●Multi-dimensional state + category switching + full sequence structure
—Business-readable state semantics at each step

SESSION-BASED RNN / TRANSFORMER FOR CONSUMER MODELING · 2018+

SYNTHESIS

Which column is your CRM in?

Translated into a business question:

Stuck in the 1959-1982 column: RFM segments, Gamma-Poisson, traditional audience packs. Sees Fate, can't see Fortune. This is where most CRM implementations actually sit.

Stuck in the 1987-2005 column: can flag churn risk, estimate P(alive). But the "living person" is the same person, with no internal change.

Reached the 2008 column: HMM and state switching. Can spot life-stage shifts, but the dimensionality is still narrow.

Reached the 2018+ column: sequence models and deep learning. Possible, at the cost of explainability.

The real move: push the CRM capability curve from 1982 to 2008, at the very least.