Six phases, arrows both ways, an outer loop

The name and the six phases

CRISP-DM stands for the Cross-Industry Standard Process for Data Mining. It is the spine of this course: every tool lesson from here on ends by naming the phase and task it belongs to, and "which phase is this?" is on every quiz, the midterm and the final.

The six phases, in the order the diagram reads:

1. Business understanding — what decision are we serving, and what would change it?
2. Data understanding — what data do we have, what is in it, and can we trust it?
3. Data preparation — the cleaned, merged, derived table the analysis will actually use.
4. Modelling — the summary, comparison or fitted line that answers the question.
5. Evaluation — does that answer actually serve the business question?
6. Deployment — putting the answer in front of the decision-maker, with a plan to keep it current.

Worked example — the Alberta question, once around

Maple & Birch's owner is weighing a second Alberta store. Business understanding produces the analytic question: does average order amount differ by province in Q1 2025, and would a gap of $100 or more change the decision? Data understanding produces a data dictionary for the exported orders file and a first look — twelve rows, four provinces, amounts from $90 to $1,320. Data preparation produces a clean table: dates parsed, amounts numeric, a per-unit column added. Modelling produces the comparison — a pivot of average amount by province: AB $1,150 · ON $498.75\cdot BC$453.33 · QC $161.67. Evaluation asks whether that answers the question: the Alberta average is built on two orders, so the gap is real in the table but cannot support a lease. Deployment would be a one-page memo saying so, with a plan to re-run the pivot when a full quarter of Alberta orders exists.

Why the arrows go both ways

CRISP-DM is drawn as a cycle, not a staircase, for two reasons.

First, adjacent phases send you backwards, and that is normal. The diagram puts two-way arrows between business understanding ↔ data understanding (looking at the data changes what question is answerable) and between data preparation ↔ modelling (building the model reveals a column you still need). Suppose the pivot above is built and the analyst realises there is no customer_type column to separate business buyers from households — that is a step back from modelling to preparation. It is not a failure of planning; it is the reason the arrow exists.

Second, evaluation can restart the whole cycle. If the evaluation says "this result would not change the decision", the honest next step is back to business understanding — refine the question — rather than on to deployment. The outer circle on the diagram is that loop: deployment is not the end, because a deployed dashboard raises new questions.

On the paper the trap is the waterfall reading. "After modelling comes evaluation" is true in the diagram's order; "once you have started modelling you never touch preparation again" is false, and the distractor for it is always a failure of planning or not allowed.

Order six shuffled phases

Cover the list and order these: deployment · data preparation · business understanding · evaluation · modelling · data understanding. Then name the two pairs with two-way arrows, and the phase the outer loop returns to.