WORKBENCH NOTE
Use this before trying to answer every question or build a model. Its purpose is to turn an unfamiliar table into a set of defensible next questions.
1. Understand the grain
- What does one row represent: a customer, order, event, account, or snapshot?
- What is the unit of observation, and can one entity appear more than once for a valid reason?
- Check whether duplicate entities or events would change the analysis.
2. Check shape and schema
- Record the number of rows and columns.
- Inspect column names, data types, and fields that look suspiciously typed.
- Ask whether dates, numeric-looking strings, IDs, and categories are represented consistently.
3. Check data quality
- Measure missingness and look for patterns, not just totals.
- Inspect duplicates, invalid values, inconsistent categories, and impossible dates or ranges.
- Document cleanup decisions and unresolved limitations.
4. Define the target or outcome
- If there is an outcome, write down its definition, prediction time, and eligible population.
- Inspect class balance or target distribution.
- Flag fields that may reveal information unavailable when a prediction would be made.
5. Inspect distributions
- For numeric fields, inspect center, spread, skew, and unusual values.
- For categorical fields, inspect frequency, rare categories, and unexpected labels.
- Treat outliers as observations to investigate, not values to delete automatically.
6. Explore relationships and time
- Compare meaningful groups, target relationships, correlations, and possible redundancy.
- If time matters, check coverage, trends, seasonality, and temporal breaks.
- Use charts to answer a question, then write the next question that the result creates.
7. Write down what you learned
- Capture anomalies, assumptions, hypotheses, and data limitations.
- Separate observation from interpretation and from a causal conclusion.
- Choose the next investigation based on evidence, not the next available chart.