Exploratory data analysis (EDA) is an approach to analysing a data set, using mostly graphical techniques, to understand its structure, spot outliers and anomalies, and decide what to model before any assumptions are fixed. Analysts use visualisation and summary statistics to see what the data can reveal beyond formal modelling or hypothesis testing.
This page is for data scientists, analysts and data leaders who start projects from raw data. It covers what EDA is, its types and steps, how it differs from statistical graphics, and what to do first.
What is exploratory data analysis?
The NIST Engineering Statistics Handbook describes EDA as an approach and philosophy for data analysis that employs a variety of techniques, mostly graphical. Its goals include maximising insight into a data set, uncovering underlying structure, detecting outliers and anomalies, and developing parsimonious models. It adds that EDA postpones the usual assumptions about what kind of model the data follow, so the data can show its own structure.
IBM says EDA originated with John Tukey in the 1970s and is used to summarise a data set's main characteristics, often with data visualisation. It sits upstream of business analytics and modelling work.
What are the types of exploratory data analysis?
IBM lists four types, based on how many variables are examined and whether the methods are graphical.
| Type | What it examines | Typical methods |
| Univariate non-graphical | One variable, to describe it and find patterns | Summary statistics |
| Univariate graphical | One variable, shown visually | Stem-and-leaf plots, histograms, box plots |
| Multivariate non-graphical | Relationships among two or more variables | Cross-tabulation, statistics |
| Multivariate graphical | Relationships among two or more variables, shown visually | Grouped bar charts, scatter plots, heat maps |
How does exploratory data analysis work?
There is no single fixed sequence, because NIST calls EDA a way of thinking rather than a collection of techniques. A common order is to identify each variable and its type, examine variables one at a time, then examine relationships between them. IBM also stresses finding missing values early, so that you can decide how to handle them before machine learning. For high-dimensional data it names clustering and dimension reduction, and it lists Python and R as widely used languages for EDA.
What is the difference between EDA and statistical graphics?
NIST says EDA is not identical to statistical graphics, although the two terms are used almost interchangeably. Statistical graphics is a set of techniques, each focused on one aspect of the data. EDA is broader and relies heavily on graphical techniques, with a few quantitative ones, to explore the data in an open-ended way.
| Question | EDA | Statistical graphics | Formal modelling or hypothesis testing |
| Starting point | The data, with model assumptions postponed | A chosen aspect of the data | A model or hypothesis fixed in advance |
| Scope | Broad, open-ended exploration | A set of plotting techniques | Tests a specific claim |
| Typical outcome | Structure, outliers, candidate models | A picture of one aspect | Support for or against the model |
Why does exploratory data analysis matter?
IBM says the main purpose of EDA is to look at data before making any assumptions. That helps teams catch missing values, odd outliers and unexpected relationships before they reach a model, a report or a decision.
How does exploratory data analysis relate to AI agents?
An agent is only as dependable as the data it reads, so EDA is a sensible check on the data an agent will rely on. Missing values and outliers found in EDA are decisions someone must own. ElixirData is XenonStack's Context OS, built for governed context across the data behind agent decisions.
What should you do first?
- List the variables in your data set and label the type of each.
- Plot each variable alone, using a histogram or box plot.
- Check for missing values and outliers, and record how you handle them.
- Examine relationships between pairs and groups of variables.
- Write down what the data suggests, and test it with formal methods afterwards.
To map one decision your agents make, and the data and policy it depends on, book a Context Workshop. See what happens in the workshop.
Frequently asked questions
What is exploratory data analysis in simple terms?
Exploratory data analysis is an approach to examining a data set, mostly with graphics, to understand its structure and spot outliers before building a model.
Who developed exploratory data analysis?
IBM says EDA originated with John Tukey in the 1970s.
What are the types of EDA?
Univariate non-graphical, univariate graphical, multivariate non-graphical and multivariate graphical.
Is EDA the same as statistical graphics?
No. NIST says the two are not identical: statistical graphics is a set of techniques, while EDA is a broader approach that relies heavily on graphical techniques.
Which tools are used for EDA?
Python and R are widely used for EDA, including for finding missing values.