Data Communication and Business Application: From Model Output to Decision
Domain scope
Data Communication and Business Application tests whether you can move from model output to a decision. It covers visualization, narrative structure, stakeholder framing, deployment, monitoring, governance, privacy, and ROI. The exam favors answers that connect a technical result to a business action and its risks.
Core rules and concepts
- Start with the decision, not the model. State the audience, the action they control, and the metric that action changes.
- Visualization should match the question: bar charts for comparisons, line charts for trends over time, scatterplots for relationships, histograms for distributions. Avoid pie charts with many slices and dual axes that invite false comparisons.
- Show uncertainty: confidence intervals, error bars, or ranges. A point estimate without uncertainty overstates confidence.
- Distinguish statistical significance from business significance. With a large sample, a tiny effect can be significant but not worth acting on.
- Deployment is a lifecycle, not a handoff: define input contracts, scoring frequency, monitoring for drift, retraining triggers, and an owner.
- Governance: data minimization, purpose limitation, access control, retention, and fairness testing for protected groups. Model cards and lineage support auditability.
- ROI should include implementation, monitoring, and false-positive/false-negative costs.
Worked example
A retention model identifies the top 10% of customers by churn probability. The business can offer a $20 credit to 5,000 customers. Historical lift shows 25% of targeted churners would stay, versus 5% without the offer. Expected saves = 5,000 × (0.25 − 0.05) = 1,000 customers. If average margin is $150, gross benefit = $150,000; offer cost = $100,000; net = $50,000 before model and operations costs. Present this as a decision: “Approve the targeted offer for the top decile; expected net benefit is $50,000 per campaign, with a break-even incremental save rate of 13.3 percentage points.” The reasoning ties model output to incremental impact, not total churn.
Traps
- Presenting model accuracy instead of lift, profit, or risk.
- Cherry-picking a favorable subgroup without disclosure.
- Ignoring base rates and control-group outcomes.
- Deploying without drift monitoring or a retraining trigger.
- Using personal data beyond its consented purpose.