Exploratory Data Analysis: A Beginner’s Project Workflow
Move from a question and data audit to qualified findings.
Part of the Data Science track: modules, projects and first roles →
By the GullyMentor team · 5 min read · Published 2026-10-18 · Updated 2026-11-06 · Sources listed at the end. Not a promise about mentorship outcomes.
Exploratory data analysis, or EDA, is a structured investigation of a dataset's quality and patterns. You do it to understand the evidence and sharpen your questions before making strong claims or building a model. A good EDA ends with a short memo, not a pile of charts.
What are the steps of an EDA?
Work in this order: frame the question, audit the data, look at one variable at a time, compare relationships, make the notebook reproducible, then write the memo. Skipping the audit is the most common reason beginner findings fall apart. The data science track places EDA right after Python and statistics.
How do you frame the investigation?
Use a fictional delivery service and ask: how does delivery time vary by zone and order period? Define the start and end timestamps. Decide which orders are included, for example completed orders only.
Record the data source, your permission to use it, the observation window and what one row means. Keep an untouched copy of the input so your cleaning decisions stay traceable.
How do you audit the data?
Inspect column types, missing values, duplicates and impossible combinations. A delivered order with no delivery time needs investigation. A negative duration may come from a parsing or time-zone problem.
A short audit table helps:
| Check | Example finding | Action |
|---|---|---|
| Missing delivery time | Some delivered orders have none | Count them, ask why, report them |
| Negative duration | End time before start time | Check time zones and parsing |
| Duplicate order id | Same order twice | Keep one, log the rule |
Create a log of every correction and exclusion. Do not remove difficult records simply because they make a chart tidier.
How do you explore one variable at a time?
Look at counts, distributions and unusual values. Compare the mean, the median and a tail measure where useful. Check how many records support each summary.
A zone with two observations should not receive the same confidence as one with two hundred. Label small groups clearly.
How do you investigate relationships?
Compare zones within similar periods. Ask whether order mix or different observation windows explain the pattern. Use plots and tables that make those comparisons visible.
Write each finding as an observation followed by a qualification. For example: "Zone B has longer observed times in this sample; order distance is unavailable, so the cause is unresolved."
How do you make the notebook reproducible?
Use explicit transformations, clear variable names and relative data paths. Restart the environment and run all cells in order. If it fails, fix it. Remove private records, credentials and accidental local paths before sharing.
What are the common mistakes?
- Cleaning data without logging what you changed.
- Dropping outliers because they look odd, not because you found the cause.
- Comparing groups without checking how many rows each has.
- Running every possible chart and keeping none of the reasoning.
- Reporting a cause when you only observed a difference.
How do you check you got it?
Hand your notebook and the raw file to a friend. They should reach the same tables by running all cells. Then ask them to read only your memo and say what you found, what you are unsure about and what you would request next. If their answer differs from yours, the memo is unclear.
How do you finish with a useful next step?
Deliver a short memo containing the question, three supported findings, the limitations and one follow-up data request. If the dataset cannot answer the original question, say what is missing.
EDA is not a ritual of producing every chart. The value lies in checking the evidence and narrowing the next investigation.
Worked example: reading one finding
Suppose the audit shows 4 of 120 delivered orders have no delivery time. First report the count. Then check whether they cluster in one zone or one day. If all four are from Zone B, a missing-data problem could explain part of any Zone B result. Write that in the memo next to the Zone B finding, and add "why are these times missing?" as your follow-up data request.
What should you build next?
Strengthen the statistics behind your qualifications with statistics for data science. Then pick a project from data science project ideas and apply this workflow to it from audit to memo. A pandas tutorial is a good companion while you practise the code.
Sources & further reading
Examples and practice recommendations are educational illustrations. Tool documentation and programme inclusions should be checked for your intended use.