Exploratory Data Analysis for CBDA: Guide Technique 3.11 Explained

Exploratory data analysis (EDA) is technique 3.11 of the IIBA Guide to Business Data Analytics: an iterative look at the data, mostly through visuals, before any formal method [1]. For the CBDA exam it matters in two domains. The guide lists it for Identify the Research Questions (20% of the exam) and describes it inside the Explore Data task of Analyze Data (16%) [2]. The skill tested is judgment, not code.

Exploratory data analysis at a glance

Exploratory data analysis is the step where you get to know the data before you trust it. The facts:

Last verified: November 2026 (exam facts on iiba.org; guide text checked section by section).

What the guide says EDA is

The guide defines EDA as "an approach used to maximize the insights gained from data" that "often uses visual analysis" before hypothesis testing, machine learning or advanced inference [1]. It is iterative, and the data is explored "without any prior assumption or bias" [1].

This is close to the classic statistics view. The NIST handbook calls Tukey's 1977 book "the seminal work in EDA" and describes the difference from classical analysis as the order of steps: classical goes data, model, analysis; EDA goes data, analysis, model [3, 4].

The guide lists nine key outcomes of EDA [1]:

• recognizing data labels, data types and the characteristics of variables in the business context;
• refining the research problem;
• finding missing and incorrect data, so the right data is sourced;
• a preliminary view of the data structure and the important variables (features or predictors);
• interdependencies, collinearity and correlations between variables;
• outliers and anomalies;
• feature engineering: derived and optimized variables for formal analysis;
• visuals that communicate initial insights;
• ancillary research questions that may matter to the business.

The guide also says domain and industry knowledge "play a critical role" in recognizing patterns and their business meaning [1]. I think this sentence is the reason EDA sits in a business analyst's guide at all.

Where EDA is used in the CBDA domains

EDA appears in Domains 1 and 3, and its limits show up in Domain 4. Unlike some chapter 3 techniques, section 3.11 has no domain-by-domain subsection, so the links come from chapter 2 and from other techniques:

For the full task picture, see Analyze Data, Domain 3 and Identify the Research Questions, Domain 1.

The nine-step EDA scheme

The guide gives a "basic step-by-step structure" while keeping the flexibility of exploration [1]. Activity diagrams, decision trees and sequence diagrams can record each step [1].

The scheme changes with the data. Images or media need dimensionality reduction and vectorization; natural language needs stop word removal and part-of-speech tagging [1].

EDA visuals

Visuals are "the primary vehicle" of EDA [1]. The choice depends on the number of variables, the data type and the step [1]:

One line here is an exam favourite. Visuals for understanding a phenomenon are "typically different" from visuals for communicating an insight [1]. EDA visuals are for the analyst. Stakeholders get business visualizations.

EDA findings

Findings need "the right packaging" [1]. The summary covers both the actions taken to clean the data and the business insights, with justifications. Analysts corroborate findings with business rule analysis, process analysis and elicitation, and may share the EDA scheme, the assumptions, each decision point and the data sourcing process [1].

The guide's own example: the $50,000 houses

Task 2.3.3 shows EDA on the Boston housing data from the UCI repository: 13 predictors and a target, the median house value in units of $1,000 (Guide, 2.3.3). Some prices look like outliers. Instead of deleting them, the analyst checks the data collection method for missed business rules. The answer: these values were "artificially set to a high value of $50,000" where prices could not be disclosed (Guide, 2.3.3). So they are excluded with a reason.

The same exploration shows that most predictors have a linear relationship with price, so ordinary least squares regression becomes a candidate model (Guide, 2.3.3). Two decisions came out of one EDA pass: data quality and modelling approach.

A worked example: why did the average order value fall?

This example is mine and the numbers are invented. An online store sees average order value (AOV) drop from about $56 in Q2 to about $53 in Q3. The research question is why.

1. Load and check. 12,000 orders, 9 columns. Region is missing in 240 rows (2%). Hold out 20% of rows as a validation set.
2. Descriptive statistics. Mean order value about $54, median $41: a long right tail, so the median is the better summary.
3. Inter-dependence. Item count and basket weight correlate almost perfectly. Keep one of them.
4. Outliers and missing data. 37 orders at $0.01. A business rule explains them: internal test orders. Exclude them, with a note. Derive missing regions from the postcode instead of imputing the most common region.
5. Visuals. A histogram by quarter shows more small orders in Q3. A boxplot by channel shows a new marketplace channel with low order values.
6. Feature engineering. Add a "first order" flag per customer.
7. Basic hypothesis. First-time buyers grew from 22% to 35% of orders. Their AOV is about $35 in both quarters; repeat buyers stay at about $62.
8. Refine the problem. Check the mix: Q2 = 0.22 × 35 + 0.78 × 62 = 7.70 + 48.36 = $56.06. Q3 = 0.35 × 35 + 0.65 × 62 = 12.25 + 40.30 = $52.55. The mix alone explains a drop of $3.51, about 6.3%. The new question: do marketplace first-timers come back?
9. Report. Two charts, the cleaning decisions (test orders, derived regions) and the refined question.

Nothing here needed a model. And the original question, "why did AOV fall", turned into a better one. That is outcome number two from the guide's list: refining the research problem [1].

Strengths and limitations

The guide lists five strengths and three limitations [1]:

My take on the first limitation: it is exactly where a business analyst earns their place in EDA. Somebody has to ask what the numbers mean for the business.

What practitioners and data say about EDA

The data says exploration and cleaning take a big share of the job. In a 2022 survey of data professionals, respondents spent about 37.75% of their time on data preparation and cleansing and 12.99% on visualization [5]. In an interview study of 18 analysts, a straightforward exploration took a few hours to a few days, and "several days or even weeks" when data was dirty or stakeholders were slow to respond [6].

Summary statistics alone are not enough, and this is old knowledge. Anscombe's four data sets share the same mean, variance and correlation (0.816) but look completely different on a chart [7]. A 2017 research paper built a dozen data sets that each produce "the same summary statistics" with very different shapes [8].

Practitioners tell what EDA catches in real work:

• A join that multiplied rows. A healthcare BI practitioner expected about 50K orders and got 547K rows: duplicate emails and a batch imported twice [9].
• Leakage found by looking. In a banking mortgage-leads project, a simple look at balances showed they already included loan amounts. The author's line: "simply looking at the data is more enlightening" than black-box models [10].
• Data that changed under the analyst. A review of a study with 30 analysts, each with at least four years of experience, describes struggles with data "versions" after collection changes [11].
• Two formulas for one metric. An analytics podcast describes two teams computing return on ad spend differently: "One says ROAS is 12. The other says it's 0.3." [12]
• An hour is often enough. A senior Python developer argues that an hour of EDA can beat "months of Dashboarding" [13].

From a CBDA angle, a candidate with about eight years as a product owner and BA describes "scenario-based questions" [14]. So expect EDA in a story, not as a definition.

Experts disagree on how to do EDA

Practitioners agree that you must look at the data first, but not on how.

1. Split before you explore. One statistics and ML writer says to create "a hygienic environment for your EDA by splitting your data" first [15]. Others on a programming forum argue EDA should be done before the split [16]. The guide sides with splitting: step 1 includes separating training and validation sets [1].

2. Automate or not. The same writer finds auto-generated reports handy, but says EDA "can't and shouldn't be automated" [15]. A team that tested two rapid EDA tools concluded: "They are not magical wands, after all." [17]

3. Exploration inflates false findings. Two statisticians warn that choices made after seeing the data inflate false positives, and that exploratory results are fragile [18]. The guide's answer is step 7 and the findings element: test a basic hypothesis, then corroborate with other techniques [1].

4. Directed or open-ended. A data scientist in biotech argues EDA "should be directed and purposeful" [19]. A popular data science textbook says "EDA is a state of mind" [20]. The guide holds both: no prior assumption, but a scheme and a research problem [1].

My view is close to the guide. Start from the research question, explore freely inside it, and write down every decision you make on the way.

How EDA differs from similar techniques

EDA is easy to confuse with its neighbours on the exam:

The neighbour I would review next is Hypothesis Formulation and Testing.

Exam traps for EDA

Check yourself: CBDA practice questions

1. During exploration, a data scientist finds a cluster of identical, unusually high prices. What should the business analyst do first? A) Delete the values. B) Replace them with the median. C) Review the data collection method for business rules that explain them. D) Build the model with and without them.

Answer: C. This is the guide's Boston housing example: the values were set to $50,000 where prices could not be disclosed (Guide, 2.3.3).

2. Which is part of the first step of the guide's EDA scheme? A) Feature engineering. B) Separating training and validation data sets. C) Reporting initial findings. D) Testing a basic hypothesis.

Answer: B. Step 1 checks dimensions and missing data and separates training and validation sets [1].

3. The team needs to present EDA results to executives. What does the guide suggest? A) Share the EDA plots as they are. B) Use explanatory analysis and visuals made for communication. C) Skip visuals. D) Share the code.

Answer: B. EDA visuals differ from visuals that communicate insight [1], and Domain 4 uses explanatory analysis (Guide, 2.4).

Conclusion: the gap between the guide's EDA and real EDA

The guide's scheme matches practice in shape, but real EDA spends more time on joins and metric definitions than the scheme suggests.

In short, learn the guide's nine steps for the exam. In real work, add one more check: are we sure every table and metric means what we think?

Related topics

• CBDA Analyze Data questions for scenario practice.
• Full mock exams in the CBDA exam simulator.

FAQ

What is exploratory data analysis in the CBDA guide?

It is technique 3.11 of the IIBA Guide to Business Data Analytics. EDA investigates, analyzes and summarizes data, mostly visually, to find patterns before formal methods such as hypothesis testing or machine learning. It is iterative and starts without prior assumptions.

What are the steps of exploratory data analysis?

The guide lists nine: check consistency and integrity, review descriptive statistics, verify inter-dependence and collinearity, treat missing data and outliers, visualize, engineer features, test a basic hypothesis, refine the research problem and report initial findings.

Who performs EDA on a CBDA project?

The guide says data exploration is primarily the data scientist's responsibility. The business analyst adds business domain knowledge, for example by checking business rules behind odd values and by judging whether results help answer the business question.

Which CBDA domains use exploratory data analysis?

The guide lists EDA as a technique for Identify the Research Questions (Domain 1) and describes it in the Explore Data task of Analyze Data (Domain 3). In Interpret and Report Results (Domain 4), analysts switch to explanatory analysis to communicate insights.

Is EDA the same as data preparation?

No. The guide says exploratory analysis is more involved than data preparation. Preparation makes data usable; EDA checks quality more deeply, looks for relationships and outliers, and guides the choice of analysis and models.

Sources

1. IIBA: Guide to Business Data Analytics, 3.11 Exploratory Data Analysis
2. IIBA: Certification in Business Data Analytics (CBDA)
3. NIST/SEMATECH e-Handbook of Statistical Methods: What is EDA?
4. NIST/SEMATECH e-Handbook of Statistical Methods: EDA vs classical analysis
5. Anaconda: 2022 State of Data Science report
6. Wongsuphasawat, Liu, Heer: Goals, process, and challenges of exploratory data analysis (2019)
7. Wikipedia: Anscombe's quartet
8. Autodesk Research: Same stats, different graphs (2017)
9. The great SQL join disaster (2025)
10. R-bloggers: Feature leakage and the case to identify it with EDA (2020)
11. The Morning Paper: Futzing and moseying (2019)
12. OWOX podcast: Facepalm moments in analytics (2025)
13. DEV Community: Exploratory data analysis, how much time and effort?
14. Techcanvass: CBDA certification journey of Aniket (2024)
15. Exploratory data analysis: what's the point? (2022)
16. Stack Overflow: EDA before or after the train/test split
17. SBC Group: Pros and cons of rapid EDA tools
18. Gelman and Loken: The garden of forking paths (2013)
19. Eric J. Ma: Exploratory data analysis isn't open-ended (2024)
20. R for Data Science: Exploratory data analysis