An independent study reference written by Dr Phuc V. Nguyen. It is not official subject material — for assessment requirements always follow your subject outline and vUWS.
CRISP-DM, the standard analytics cycle
CRISP-DM (Cross Industry Standard Process for Data Mining) describes analytics work as six phases: business understanding, data understanding, data preparation, modelling, evaluation and deployment. Two features matter more than the list. The arrows run both ways between business and data understanding, and between preparation and modelling, because each teaches you something about the other. And evaluation is a decision point rather than a checkpoint. It sends the work forward to deployment, back to business understanding for reframing, or into another iteration. It is method-neutral, which is why it has outlived the tools it was written for.
Why it matters
Picture a loop with two tight inner loops inside it. You learn what the business wants, you look at the data, and each look changes the question. Then you shape the data, try a model, and each attempt sends you back to reshape something. At the end you do not ask "is the model good", you ask "does this meet what we set out to do, and what did we miss". Then you choose: put it into use, run the loop again with better questions, or reframe the problem.
During data understanding a team discovers the warehouse holds no field identifying which customers received the offer being studied. What does CRISP-DM indicate should happen next?
Worked examples
A subscription business wants to know which customers will cancel. Halfway through data understanding, the team finds the warehouse never recorded which customers were offered a retention discount.
CRISP-DM sends the team back here, not forward. Without a treatment flag the data cannot separate customers who stayed because of the discount from those who would have stayed anyway, so the original question cannot be answered honestly. The team returns to business understanding and reframes: predict cancellation risk without claiming what caused it. Recording the flag from now is necessary but not sufficient, because agents chose whom to approach, so offered and unoffered customers differ in ways no flag records. Answering it causally needs a randomised offer, a holdout group, or identification assumptions stated in the open. The backtrack cost two days. Building on the gap would have cost a year of confident, wrong conclusions.
A churn model reports unusually strong accuracy. In evaluation, the team notices one of its strongest inputs is a case-notes field that is only populated after a customer has phoned to cancel.
This is leakage, and it is evaluation doing the job the modelling phase cannot. Modelling asks a technical question, whether the model fits and generalises on held-out data, and a leaked field passes that test easily because it genuinely predicts the outcome in historical data. Evaluation asks a different question: does this meet the business objective, and has anything been overlooked. At the moment a prediction is actually needed, the field is empty, so the model would be useless in production. The field is removed and modelling is repeated.
Common mistakes
- ✗CRISP-DM is a linear pipeline you walk through once. It is drawn as a cycle with explicit backward arrows, and the most common backtrack is from data understanding to business understanding, when the data turns out not to support the question as asked.
- ✗Evaluation means checking accuracy. Accuracy and error are assessed in the modelling phase. Evaluation is a business review: does the result meet the objective that was agreed, and was anything important overlooked, including leakage, an unusable input or a group the model was never tested on.
- ✗Deployment means putting a model into software. Deployment is whatever makes the result usable by the people who need it. That can be a scheduled report, a rule handed to a team, a threshold written into an existing process, or an automated service. The test is that someone can act on it without the analyst present.
- ✗CRISP-DM is obsolete because it predates modern machine learning. It is method-neutral, so the phases apply unchanged to a regression or a neural network. Its genuine gap is the life of a model after go-live, meaning monitoring, drift and governance, which later work extends rather than replaces.
Revision bullets
- •Six phases: business understanding, data understanding, preparation, modelling, evaluation, deployment
- •Two-way arrows: business and data understanding, preparation and modelling
- •Evaluation is the decision point: deploy, iterate, or go back and reframe
- •Modelling asks "does it fit", evaluation asks "does it meet the objective and what did we miss"
- •Data preparation is usually the longest phase
- •Method-neutral, but silent on life after go-live
Quick check
During data understanding a team discovers the warehouse holds no field identifying which customers received the offer being studied. What does CRISP-DM indicate should happen next?
A model scores very well on held-out data, but one strong input is only populated after the event being predicted has begun. Which phase is doing its job by catching this, and why?
Connected topics
More in How Analytics Gets Built
Sources
- Chapman et al. (2000)Chapman, P., Clinton, J., Kerber, R., Khabaza, T., Reinartz, T., Shearer, C., & Wirth, R. "CRISP-DM 1.0: Step-by-Step Data Mining Guide." CRISP-DM consortium, 2000.The reference document defining the six phases, their generic tasks and their outputs.
- Shearer (2000)Shearer, C. "The CRISP-DM Model: The New Blueprint for Data Mining." Journal of Data Warehousing, 5(4), 13-22, 2000.A compact account of the cycle and the reasoning behind the phase boundaries.
- Martinez-Plumed, F., Contreras-Ochando, L., Ferri, C., Hernandez-Orallo, J., Kull, M., Lachiche, N., Ramirez-Quintana, M. J., & Flach, P. "CRISP-DM Twenty Years Later: From Data Mining Processes to Data Science Trajectories." IEEE Transactions on Knowledge and Data Engineering, 33(8), 3048-3061, 2021.Reviews where the original cycle still holds and where modern practice departs from it, including continuous operation.