IS6400 Chap.1 Business Analytics Workflow
Business Analytics Workflow
Begin with the decision
The course places business understanding, data understanding and preparation before model building. That order prevents a team from optimising a score for a target that does not match an operational choice.
Define who acts, what is decided, when it is decided, what one row represents and what later outcome will test success.
Backtracking is part of the method
Testing and evaluation may expose an earlier mistake. Leakage sends the project back to preparation; unstable segment results may send it back to data understanding; a technically strong score that changes no action sends it back to business understanding.
Deployment therefore includes a threshold, owner, action, monitoring measure and exception route.
Write the information boundary
Place the decision timestamp on a timeline and classify every candidate field as available before, at or after that point. Fields created after the outcome are leakage even when they are stored in the same convenient table.
Historical aggregates need their own cutoff rule so a later transaction cannot enter an earlier prediction. Keep the rule with the feature definition, because a clean train-test split cannot repair a feature that already contains future information.
Evaluate the action, not only the score
Choose a baseline that represents the current process, then test the candidate on later or otherwise independent rows.
Translate its errors into workload, missed opportunities, waste or another consequence the organisation recognises. A threshold determines how many cases receive action, so report performance at that threshold and check capacity.
If the model changes no decision, the project needs a new action rule or a return to business understanding rather than another tuning cycle.
Make deployment reversible
Name the operational owner, monitoring frequency, inputs that may drift and outcome that will arrive later. Define an exception path for cases outside the training range and an escalation trigger for workload or harm.
Save the model version, preprocessing rules, threshold and baseline comparison. These details let the team distinguish a changed population from a coding failure and let it pause or roll back the system without reconstructing the original analysis from memory.
What this chapter covers
- 01
Business understanding and an operational owner
- 02
One row and one prediction horizon
- 03
Data understanding before feature construction
- 04
Preparation without future leakage
- 05
Testing against a baseline
- 06
Deployment, monitoring and backtracking
Worked example · free
Define a replenishment decision before modelling
- 2Specify the decision unit and horizon. The marks shown here are not an official University mark allocation.
- 3Name fields that would leak future information.
- 3Define the baseline, action and monitoring trigger.
Key terms
- Decision Unit
- The entity and moment to which one prediction and one action apply.
- Target Horizon
- The future interval between the decision point and the outcome being predicted.
- Data Leakage
- Information entering a model that would not be available when the live decision is made.
- Baseline Model
- A simple reference rule that a more complex candidate must beat under the same evaluation.
- Monitoring Trigger
- A stated condition that starts investigation, retraining or escalation after deployment.
Business Analytics Workflow FAQ
Why should the business question come before the model?
Because the question fixes the row, target, horizon, costly error and action. Without those choices a model can optimise a technically valid score that cannot be used, or can learn from information that arrives after the real decision.
Does a weak test result always mean the algorithm is too simple?
No. It may reveal a vague target, censored outcome, missing segment, bad join or time leakage. Return to the earliest stage contradicted by the evidence before increasing model complexity.
What turns a prediction into a deployed analytics process?
A deployment identifies the operational owner, operating threshold, action, exception route, monitoring measure and escalation trigger. A score delivered without those pieces is an analytical output, not a controlled business capability.
Exam move
Take one familiar business request and write seven lines: operational owner, action, row, target, horizon, costly error and baseline. Audit every feature against the decision timestamp. Then sketch the six process stages and place each possible failure at the earliest stage able to repair it. Finish by naming what would be monitored after release and which change would trigger retraining rather than a threshold adjustment.
Rehearse one failure at every stage: a vague action, an ambiguous row, a censored target, a leaking field, a mismatched metric and an unowned deployment. For each, point to the earliest repair and explain why later model tuning cannot solve it. End with a one-minute spoken summary that connects the baseline comparison to a reversible operational decision.