Exploratory Data Analysis (EDA): A Step-by-Step Guide in Python

Exploratory data analysis (EDA) is the first look an analyst takes at a dataset before building reports, models, or conclusions. It means checking what the data contains, how values are distributed, what is missing or strange, and how variables relate to each other, mostly with simple summaries and charts. EDA matters because it catches problems early: a column full of placeholder values, a date range that stops in the middle of a month, or a relationship that only exists because of a few outliers. Skipping it is how confident analyses end up being wrong.

This guide walks through a practical EDA process in Python with pandas, the questions to ask at each step, and how to turn what you find into decisions.

Where EDA comes from and why it still matters

The statistician John Tukey popularized the idea in the 1970s: look at the data first, with open questions, before testing a fixed hypothesis. The NIST Engineering Statistics Handbook still describes EDA this way, as an approach that relies heavily on graphics to reveal structure, spot outliers, and check assumptions.

A classic reminder of why this matters is Anscombe's quartet: four small datasets with nearly identical means, variances, and correlations that look completely different when plotted. Summary numbers alone can hide the story. EDA makes you look.

Step 1: Start with the business question

EDA is open-ended, but it should not be aimless. Write down what the data is supposed to help decide, for example "why did repeat purchases drop?" That question tells you which columns matter most and when you have explored enough.

Step 2: Get to know the shape of the data

import pandas as pd
df = pd.read_csv("orders.csv", parse_dates=["order_date"])
df.shape        # rows and columns
df.head()       # first rows
df.dtypes       # data type of each column
df.nunique()    # distinct values per column

import pandas as pd
df = pd.read_csv("orders.csv", parse_dates=["order_date"])
df.shape        # rows and columns
df.head()       # first rows
df.dtypes       # data type of each column
df.nunique()    # distinct values per column

import pandas as pd
df = pd.read_csv("orders.csv", parse_dates=["order_date"])
df.shape        # rows and columns
df.head()       # first rows
df.dtypes       # data type of each column
df.nunique()    # distinct values per column

import pandas as pd
df = pd.read_csv("orders.csv", parse_dates=["order_date"])
df.shape        # rows and columns
df.head()       # first rows
df.dtypes       # data type of each column
df.nunique()    # distinct values per column

Questions to answer here: what does one row represent (an order, a customer, an order line)? Are numbers stored as numbers and dates as dates? Does a column that should be unique, like order_id, actually have one value per row?

Step 3: Check missing and suspicious values

df.isna().sum().sort_values(ascending=False)
df["country"].value_counts(dropna=False).head(20)

df.isna().sum().sort_values(ascending=False)
df["country"].value_counts(dropna=False).head(20)

df.isna().sum().sort_values(ascending=False)
df["country"].value_counts(dropna=False).head(20)

df.isna().sum().sort_values(ascending=False)
df["country"].value_counts(dropna=False).head(20)

Missing values are not always NaN. Look for placeholders like 0, -1, "N/A", "unknown", or 1900-01-01 dates. The pandas guide to missing data explains how pandas represents and handles them. For each gap, decide whether to fill it, drop it, or keep it and flag it, and write down why. If the data needs serious fixing, our data cleaning guide with pandas goes deeper.

Step 4: Summarize each variable on its own

df.describe()                      # numeric columns
df.describe(include="object")      # text columns
df["amount"].hist(bins=50)

df.describe()                      # numeric columns
df.describe(include="object")      # text columns
df["amount"].hist(bins=50)

df.describe()                      # numeric columns
df.describe(include="object")      # text columns
df["amount"].hist(bins=50)

df.describe()                      # numeric columns
df.describe(include="object")      # text columns
df["amount"].hist(bins=50)

describe() gives count, mean, standard deviation, minimum, quartiles, and maximum. Compare the mean with the median (the 50% value): if the mean is much higher, a few large values are pulling it up, which is typical for revenue and order amounts. A histogram or box plot shows the shape directly: is it skewed, does it have two peaks, are there impossible values like negative quantities?

For categories, look at counts. A "country" column with "US," "USA," and "United States" needs standardizing before any grouping.

Step 5: Look for outliers, then investigate them

Outliers are not automatically errors. A $50,000 order could be a typo or your biggest corporate client. Flag extreme values with a simple rule, such as values beyond 1.5 times the interquartile range, then look at the actual rows before deciding anything. Removing outliers just to make a chart look cleaner changes the answer to the business question.

Step 6: Explore relationships between variables

df.groupby("channel")["amount"].agg(["count", "mean", "median"])
df[["amount", "items", "discount"]].corr()
df.set_index("order_date")["amount"].resample("W").sum().plot()

df.groupby("channel")["amount"].agg(["count", "mean", "median"])
df[["amount", "items", "discount"]].corr()
df.set_index("order_date")["amount"].resample("W").sum().plot()

df.groupby("channel")["amount"].agg(["count", "mean", "median"])
df[["amount", "items", "discount"]].corr()
df.set_index("order_date")["amount"].resample("W").sum().plot()

df.groupby("channel")["amount"].agg(["count", "mean", "median"])
df[["amount", "items", "discount"]].corr()
df.set_index("order_date")["amount"].resample("W").sum().plot()

Group by the categories that matter to the question, check correlations between numeric columns, and plot key metrics over time. Scatter plots help you see whether a correlation is real or driven by a few points. Libraries like seaborn make distribution and relationship plots quick to produce. If you are unsure which chart fits each question, see our guide to types of charts and graphs.

Remember that correlation is not causation. EDA finds patterns worth testing; it does not prove why they happen.

Step 7: Check time coverage and consistency

Many surprising "drops" are data problems: a tracking change, a missing day, or a partial current month. Plot row counts per day or week. If the count falls off a cliff, ask whether the business changed or the pipeline did.

Step 8: Write down what you found

End EDA with a short list: data quality issues and how you handled them, the main patterns, open questions, and what analysis comes next. This turns exploration into something your manager or team can act on, and it keeps you honest about assumptions.

Common EDA mistakes

• Jumping to models or dashboards before understanding the data.

• Trusting averages without looking at distributions.

• Dropping missing rows by default, which can bias results if the missing data is not random.

• Exploring forever. Stop when you can answer the business question or know exactly what data is missing.

• Not saving your work. Keep EDA in a notebook so others can rerun it.

Using AI assistants during EDA

AI assistants can speed up EDA by suggesting pandas code, summarizing columns, or proposing charts. Treat their output as a draft: run the code yourself, check row counts, and confirm any claim against the data. Statistical basics still matter here, and our overview of statistics for data analysts covers the concepts you will use most.

Frequently asked questions

Is EDA the same as data cleaning?

They overlap. EDA is about understanding the data; cleaning is about fixing it. In practice you alternate: exploring reveals problems, cleaning fixes them, and you explore again.

Can I do EDA in Excel or SQL instead of Python?

Yes. Pivot tables, GROUP BY queries, and simple charts all count as EDA. Python becomes more useful as datasets get larger and you want repeatable steps.

How long should EDA take?

It depends on the data and the stakes. For a quick question on a familiar dataset, an hour may be enough. For a new dataset that feeds an important decision, plan for more time.

Practice EDA with live guidance

If you want to learn EDA on real business cases with feedback from practitioners, Coderhouse offers live online courses with one class per week:

• Data Analytics Course: 11 weeks for career switchers and data-driven professionals, including a week on Python and exploratory analysis, plus SQL, visualization, statistics, and AI-assisted analytics.

• AI Engineering Course: 8 weeks for experienced developers moving into AI engineering, covering retrieval, RAG, agents, evaluation, and deployment.