Three numbers and a snapshot date

R, recency: days between the customer's last order and the snapshot date. Smaller is better, which is why R is the one dimension scored in reverse later on. Two choices: which date is "last order" (placed, approved, or delivered) and which date is "now". On Olist the placed timestamp is the honest one, because delivered dates depend on the carrier rather than on the customer.

F, frequency: the number of orders. Choices: all time or a window (the last 12 months is common, and it changes who counts as frequent), and whether cancelled, unavailable and undelivered orders count. Here only delivered orders count, which drops 2,963 of 99,441.

M, monetary: what the customer is worth in money. Choices: total spend or average basket (total rewards frequency twice, since F is already counted; average basket does not), and which column is "spend". Olist has both an items table (price plus freight per line) and a payments table (what was actually charged, including vouchers and instalments). The payments table is used here because it is the money that moved; 2,961 orders were paid with more than one payment and they are summed.

None of these is the right answer everywhere. What matters is that each is written down in the code, once, so that when a second analyst gets a different table the difference is a diff and not an argument.

RFM is direct-mail technology. The version most people mean is Arthur Hughes's from Strategic Database Marketing (1994): sort the customer file by recency into fifths, sort each fifth by frequency into fifths, each of those by monetary, and you have 125 cells; then mail a random tenth of every cell, read the response rate per cell, and mail the rest of only the cells that paid. The scoring was never the point. The test mailing was, and session 5 is about the modern version of that step.

Many tutorials say RFM has been "in use since 1961". That date travels from page to page without a citation, and a search for its origin during this build found none. The direct-mail roots are real and usually dated to the 1960s catalogue business, but there is no founding paper and no single year to defend. Hughes 1994 is the citable anchor. When a fact has been copied more often than it has been checked, teach the uncertainty rather than the number.

The Olist export's last timestamp is 17 October 2018, so the obvious snapshot is df["order_purchase_timestamp"].max(). It is wrong by 47 days. Only 20 orders are dated September or October and not one of them was delivered: the export stopped at the end of August and a handful of stragglers leaked in. Measured from 17 October, every customer looks a month and a half staler than the data can support, and the whole recency column shifts.

The rule that survives: plot orders per month, find the last complete one, snapshot on the first day after it. Here that is 1 September 2018. Write it as a literal in the code, not as a computed max, so that re-running next quarter is a deliberate edit rather than a silent drift.

This is the same frozen-history discipline the retention course applies to cohort tables, and it matters here for the same reason: a segment view recomputed against a moving "today" quietly restates the past every time it runs.

Olist's orders table carries customer_id, and there are 99,441 of them for 99,441 orders: it is an order-side key, one per purchase. The customers table maps it to customer_unique_id, of which there are 96,096. Group by the first and every customer bought exactly once, by construction, and frequency is meaningless. Group by the second and 2,801 customers turn out to have bought more than once.

Every order system has this pair somewhere: a session or checkout id and a person id. Frequency is only defined against the person, and checking which one you have is the first line of any RFM build, before any date is chosen.

Put the knife at the mean of each column, the most common tutorial choice, and read what the eight groups contain on 93,358 real customers. The mean order count is 1.033, so "frequent" means two or more orders and only 3.0 percent of customers clear it. The mean spend is R$ 165 against a median of R$ 108, so the spend column is skewed right and only 29.3 percent are "big". The mean recency is 239.5 days and 54.8 percent are more recent than that.

The result: High value (111) is 1.28 percent of customers and holds 3.1 percent of revenue. The biggest group is General develop (100), recent one-time buyers with a small basket, at 37.9 percent. And the group holding the most revenue is not the corner with three ones at all: Key develop (101), recent single big baskets, is 15.1 percent of customers and 32.3 percent of revenue.

None of this is a flaw in the taxonomy. It is what a marketplace where 97 percent buy once looks like through it, and it is the reason the eight names should be read as coordinates rather than as verdicts. Session 2 moves the knife and watches every one of these shares change.