By a segmentation model I mean tiers by value: every customer ranked by the margin they bring, and the ranking cut into a few groups that each get their own treatment. Clusters found by an algorithm on customer behaviour are a different job, and a longer one. They arrive with no meaning attached, so someone has to name each cluster and decide what it gets; a value tier comes with that built in.
Where the days go
Let’s take an invented online coffee shop: 45,000 customers who ordered at least once in the past year, orders in the shop platform, product costs in the accounts. My estimate for each step, in working days:
The model is 1 to 2 of the 6 to 14 days; getting the data together is the widest step
The same data as a table
| Step | Value already brought | Value expected |
|---|---|---|
| Agree what the tiers are for and what counts as value | 1 to 2 | 1 to 2 |
| Get orders, refunds and product costs into one table | 2 to 5 | 2 to 5 |
| Work out a value for every customer | 0.5 to 1 | 2 to 4 |
| Cut the ranking into tiers | 0.5 to 1 | 0.5 to 1 |
| Check the tiers hold, on past quarters | 1 to 2 | 1 to 2 |
| Put the tiers where they are used, and set up the monthly refresh | 1 to 3 | 1 to 3 |
| Total | 6 to 14 | 7.5 to 17 |
The model itself, a value for every customer and the cut into tiers, is 1 to 2 of the 6 to 14 days. The widest step is getting the data together: 2 days if orders, refunds and costs export cleanly, 5 if product costs have to be matched to orders by hand.
Agreeing what the tiers are for is little analysis and a lot of waiting, because it is a decision, and it belongs to whoever will act on the tiers. Until it is made, value has no definition. Margin or revenue, the past year or the next one, refunds in or out: each choice gives a different ranking.
Value already brought, or value expected
The one choice that changes the plan is which value you rank by. Value already brought is arithmetic on the order table, each customer’s margin over the past 12 months, and takes a day at most. Value expected is a forecast for each customer, and a forecast has to be tested on months it never saw before anyone acts on it. That is where the extra 1.5 to 3 days go.
One fork in the plan: ranking by value expected adds 1.5 to 3 days
The forecast needs no more data than the arithmetic. Fader, Hardie and Lee built a model that forecasts each customer’s value from three numbers alone: when they last bought, how many times, and how much on average. It is for customers who buy when they please rather than on a subscription, and all three numbers come from the order table. They tested it the way any forecast should be tested, on 2,357 customers of the online music site CDNOW: fitted on 39 weeks of purchases, checked against the 39 weeks that followed. A check like that, they noted, is missing from virtually all lifetime-value papers.
Why spend the extra days? Because value already brought puts some of the wrong customers at the top. The same paper found what its authors call the increasing-frequency paradox: among customers who have not bought for a long time, the ones who used to buy most often are worth less than the ones who bought rarely, because a long silence from a frequent buyer most likely means they have gone. In the paper’s example, customer A, a slow buyer probably still active, is expected to make 4.6 more purchases, discounted to today; customer B, who bought often and then stopped, 1.9. Ranked by what each has already brought, and with purchases of the same size, B comes first.
Ranked by what each has brought, B comes first; by what each is expected to bring, A does
For a subscription business, the expected value comes from churn measured by plan and by how long customers have stayed; the formula and its corrections are on customer lifetime value.
The check takes 1 to 2 days, not a quarter
Tiers are only usable if a customer’s tier does not reshuffle every month, and you do not have to wait a quarter to find out. Rank the customers as of past dates and see who moved. Let’s back-test the invented shop’s top tier: the 4,500 customers ranked highest on 1 January by their margin over the previous 12 months.
Re-ranked on 1 April, 3,825 of them were still in the top tier, 85%. That looks stable, and part of it is arithmetic: the two 12-month windows share 9 months. The better question is who stopped buying. 765 of January’s top tier bought nothing from January to March, and 652 of those were still in the top tier in April, carried there by purchases made up to a year earlier. They are 17% of the 3,825 still on top, and they are the customers the paradox above is about.
85% of January’s top tier was still on top in April, 652 of them without a purchase all quarter
The same data as a table
| On 1 April | Customers |
|---|---|
| Still in the top tier, bought in the quarter | 3,173 |
| Still in the top tier, bought nothing all quarter | 652 |
| Moved to the middle tier | 608 |
| Moved to the rest | 67 |
| January’s top tier | 4,500 |
How stable is stable enough depends on what the tiers decide. I would want a top tier that holds for a year before giving each of its customers an account manager; a monthly offer list can live with far more movement.
What makes it take longer
Each of these lands in the first two steps, before any model runs:
- orders in more than one system, with no customer ID the systems share;
- no cost per product, so margin has to be estimated before it can be ranked;
- refunds and returns kept apart from the orders they reverse;
- one customer holding several accounts, or several people sharing one;
- nobody free to decide what the tiers are for.
How FinDaS does it
Tiers by value are one deliverable, priced as a sprint: from €12k, fixed in a one-page proposal before work starts. AI agents build the ranking and the back-test, and a review round checks them. The memo that says what each tier gets goes to a panel of reviewer agents, and then I sign it. Routine monthly refreshes of the tiers run on the agents’ checks. The steps are on how it works.
Sources
- Peter S. Fader, Bruce G. S. Hardie and Ka Lok Lee, RFM and CLV: Using Iso-value Curves for Customer Base Analysis, Journal of Marketing Research 42(4), November 2005, pp. 415-430. The figures here are from the authors’ version, on Bruce Hardie’s site.