ECON7880 Chap.1 Analytics Questions and the CRISP-DM Cycle
Analytics Questions and the CRISP-DM Cycle
Start with the decision
This chapter turns an organisational request into a population, prediction time, future target, feature cutoff and action.
It uses the CRISP-DM cycle to show why modelling is only one stage and why evaluation can send a team back to the business question or data preparation.
Keep prediction and explanation separate
A useful predictive association may support ranking without proving that changing the predictor will change the outcome.
The worked case uses repeated customer rows to expose identity and time leakage before any model is selected.
Define one row before choosing a target
A row might represent a customer, order, customer-month or transaction. That choice controls what can be predicted and how observations relate to one another.
Repeated rows for the same entity are not independent merely because they occupy different lines in a table. Write the row unit, prediction moment and target window together, then verify that the proposed action happens after the prediction and before the target becomes known.
Make the feature boundary operational
The feature cutoff is a time rule rather than a list assembled after seeing the data.
Information available only after cancellation, default or return belongs outside the predictor table even when it produces a large score improvement. Apply the cutoff to every join and aggregation.
A historical count is legal only if its contributing events occurred before the case was scored, and the notebook should make that condition inspectable.
Choose evidence for the intended deployment
A future-facing time split tests whether relationships survive changing conditions. A grouped split tests whether the method transfers beyond repeated entities already seen. Some projects need both.
The test design should imitate the next cases the organisation expects rather than maximise convenience.
Evaluation then connects the score to a workload, threshold and error cost, completing the move from model output to a responsible decision.
Return to the cycle when assumptions fail
If a target is ambiguous, a feature arrives too late or performance collapses in a later period, the answer is not automatically a more complex algorithm.
Return to business or data understanding, refine the question and rebuild the evidence boundary. CRISP-DM is iterative precisely because failures during modelling or evaluation reveal information about earlier choices. Record that return as part of the analysis rather than hiding it as a discarded attempt.
What this chapter covers
- 01
Business questions and decision owners
- 02
Prediction time and target windows
- 03
Feature cutoffs and leakage
- 04
CRISP-DM stages
- 05
Deployment and monitoring
- 06
Prediction versus causal explanation
Worked example · free
Repair a leaking churn design
- 3Name identity and time leakage.
- 3Choose a future-facing split.
- 2Tie the threshold to action capacity.
Key terms
- Prediction Time
- The moment at which a deployed model must produce its score using only information then available.
- Feature Cutoff
- The latest permitted time for information entering predictors.
- Deployment Drift
- A change in data or decisions that weakens the relationship learned during training.
Analytics Questions and the CRISP-DM Cycle FAQ
Why begin with a decision rather than a model?
The decision fixes the population, target, timing, action and costly error. Without those choices, an impressive score may answer a different question from the one the organisation can act on.
When is a random split unsafe?
It is unsafe when time, groups or repeated entities let related information cross between training and test data. The split should imitate how future cases will arrive.
Does prediction show causation?
No. Prediction can exploit association for ranking, while an intervention requires evidence that changing a factor changes the outcome. State which claim the task actually needs, which action will follow and what additional design would support a causal conclusion. Keep prediction and intervention claims in separate sentences, with distinct evidence requirements and limitations.
Exam move
Take one business request and write five lines: population, prediction time, target, feature cutoff and action. Only then choose a split and metric. Test the design by asking whether a new analyst could reproduce the eligible row and feature set from timestamps alone. Add a baseline, an operating capacity and the costly error.
Finally, change the deployment scenario from future cases to unseen customers and decide whether the split must change. Repeat with a second context until each line is specific enough to audit and every evaluation choice follows from the intended use. Build a small leakage checklist covering post-outcome fields, repeated entities, global preparation and incomplete target windows.
For each failure, name the CRISP-DM stage that should be revisited and the evidence needed before modelling resumes. End by presenting the plan in sixty seconds without naming an algorithm; if the decision and validation design are still clear, the framing is doing its job.
Create a final one-page audit linking each business statement to its data field, timestamp, metric and owner, then identify any assumption that remains unverifiable before deployment.
Working through Analytics Questions and the CRISP-DM Cycle in ECON7880? Sia is AskSia’s AI Economics tutor — ask any ECON7880 Analytics Questions and the CRISP-DM Cycle question and get a clear, step-by-step explanation grounded in how ECON7880 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
