Memory hook: A pattern is evidence to investigate, not proof of cause.
Must remember
Grouping combines categories into meaningful sets; binning organizes numerical or time values into ranges; clustering seeks groups with similar characteristics. Choose bins thoughtfully because changing boundaries can alter the apparent distribution. Explain the population, time range and filters behind a chart.
Reference lines mark targets or summary values. Error bars communicate uncertainty or variability under the chosen definition. Forecasting extrapolates a modeled pattern and needs enough suitable history; structural changes can invalidate it. An anomaly is an unusual observation relative to a model, not automatically a data error or fraud.
AI visuals such as key influencers and decomposition trees help explore relationships and contributions. Use the Analyze feature to investigate changes, but validate suggested explanations against domain knowledge and data quality. Correlation does not establish a causal intervention would have the predicted effect.
Copilot summaries of a semantic model depend on meaningful names, definitions and trustworthy measures. Preserve privacy and permissions, and check eligibility/capacity requirements before promising a feature to stakeholders. Good analysis distinguishes observation, hypothesis and verified conclusion.
Recall drill: sales rose after a campaign, but prices and seasonal demand changed at the same time. State what the chart proves, what it cannot prove, and what comparison or experiment would improve the conclusion.
Choose under exam pressure
| Requirement | Choice and reason |
|---|---|
| Show spread and uncertainty | Appropriate error bars with a clear definition. |
| Explore drivers of an outcome | Key influencers/decomposition with validation. |
| One extreme transaction | Investigate context and quality before deleting it. |
Traps
- A forecast is not a guaranteed future value.
- A correlation in a filtered dataset may not hold in the full population.
Active recall
1. Grouping versus binning?
Combining categories versus creating numerical/time ranges.
2. What does an anomaly indicate?
A deviation from an expected pattern that merits investigation.
3. Why label uncertainty?
Readers otherwise may treat an estimate as an exact fact.
4. Why can a forecast fail after a policy change?
The historical relationship may no longer describe the new regime.
5. Does a key-influencer result prove causation?
No; it identifies associations under its data/model assumptions.