Data Analysis Skill

Answer data questions ranging from quick metric lookups to full investigative analyses. Adapt depth and rigor to the complexity of the question.

Workflow

Step 1: Classify the Request

Determine the analysis tier before starting work:

Step 2: Understand the Question

Before touching any data, confirm:

1. What is the core question? Restate it in your own words.
2. What metric(s) matter? Identify the KPIs or measures involved.
3. What dimensions are relevant? Time periods, segments, geographies, products.
4. What granularity? Daily, weekly, monthly, per-user, per-transaction.
5. Who is the audience? Executive summary vs. analyst deep-dive.
6. What action will this inform? Understanding the decision context shapes the analysis.

If any of these are ambiguous, ask clarifying questions before proceeding.

Step 3: Gather Data

Depending on what is available:

• SQL against a warehouse: Write and run queries. Use CTEs for readability. Always include row counts and sanity checks.
• CSV/Excel files: Load with pandas. Immediately check shape, dtypes, nulls, and head/tail.
• API responses: Parse JSON, flatten nested structures, validate completeness.
• Provided tables or numbers: Verify units, time periods, and whether values are cumulative or incremental.

Always perform a data quality check:

# Standard quality check
print(f"Shape: {df.shape}")
print(f"Date range: {df['date'].min()} to {df['date'].max()}")
print(f"Null counts:\n{df.isnull().sum()}")
print(f"Duplicates: {df.duplicated().sum()}")

Step 4: Analyze

Follow these principles:

• Start broad, then narrow. Look at the overall trend before drilling into segments.
• Compare to a baseline. Week-over-week, month-over-month, year-over-year, or plan/budget.
• Decompose changes. Break a top-line change into contributing factors (mix shift, volume, rate).
• Quantify impact. Every insight should have a number attached: "Segment X declined 12%, contributing 4pp of the total 7pp decline."
• Check for confounders. Seasonality, holidays, product launches, data pipeline issues.
• Validate with a second approach. If you computed a metric one way, verify with an alternative calculation.

Step 5: Structure the Output

For Lookups

**[Metric Name]**: [Value] ([comparison to baseline])
Context: [One sentence of relevant context]

For Explorations and Comparisons

## Summary
[2-3 sentence answer to the question]

## Key Findings
1. [Finding with quantification]
2. [Finding with quantification]
3. [Finding with quantification]

## Supporting Data
[Table or chart]

## Caveats
[Any data quality issues, missing data, or assumptions]

For Deep Analyses

## Executive Summary
[3-5 sentences answering the question with key numbers]

## Methodology
[How you approached the analysis, what data was used, time period]

## Findings
### Finding 1: [Title]
[Detail with supporting data]

### Finding 2: [Title]
[Detail with supporting data]

## Recommendations
[What actions the data supports]

## Appendix
[Detailed tables, query logic, caveats]

Examples

Example 1: Quick Lookup

User: "What was our conversion rate last week?" Response: Pull the metric, compare to prior week and same week last year, note if it is within normal range.

Example 2: Trend Investigation

User: "Why did revenue drop in March?" Response: Decompose revenue into volume and average order value. Break down by channel, geography, and product line. Identify the largest contributing factors. Check for external events (holidays, outages, competitor actions).

Example 3: Segment Comparison

User: "How do enterprise customers compare to SMB in retention?" Response: Compute retention curves for both segments. Compare at 30/60/90/180/365 day marks. Break down by cohort vintage. Test if the difference is statistically significant.

Best Practices

1. Always show your work. Include the queries or code that produced each number so results can be reproduced.
2. Lead with the answer. State the conclusion first, then provide supporting evidence. Do not make stakeholders read through methodology to find the answer.
3. Use appropriate precision. Report "$4.2M" not "$4,217,893.42" for executive audiences. Keep full precision in appendix tables.
4. Flag uncertainty. If data is incomplete or methodology has limitations, say so explicitly.
5. Distinguish correlation from causation. Be precise about what the data shows vs. what it implies.
6. Include time context. Always state the date range analyzed and when the data was last refreshed.
7. Normalize when comparing. Use per-capita, per-session, or indexed values when comparing groups of different sizes.
8. Provide actionable next steps. End with what should happen next, whether that is a deeper investigation, a decision, or monitoring.