Technical visualizations are technique 3.18 of the IIBA Guide to Business Data Analytics: charts that data scientists use to understand data, choose an approach and check model assumptions [1]. They "may not be useful" for business stakeholders [1]. For the CBDA exam they sit in Analyze Data, 16% of the 75 questions [2]. The analyst usually does not build these plots. The analyst reads and translates them, and that translation is what I would study first.
Technical visualizations at a glance
Technical visualizations are the instruments of the analysis, not its output. The facts:
| Item | Value |
|---|---|
| Guide section | 3.18, chapter 3 (Techniques) [1] |
| Purpose | Help data scientists evolve their analysis and deepen the team's understanding [1] |
| Listed in | 2.3.6, Select Techniques for Analyze Data (no BABOK reference) (Guide, 2.3.6) |
| Also used in | 2.4.4, Derive Insights from Data, with an error residue graph example (Guide, 2.4.4) |
| Elements | Eight common visuals: autocorrelation, Box-Cox, contour, correlation map, PCA, residual, ROC, silhouette [1] |
| Analyst's role | Be comfortable with the common visuals, understand the analysis, then "simplify, translate, and communicate" [1] |
| Exam | 75 multiple-choice, scenario-based questions in 2 hours; Analyze Data is 16%, Interpret and Report Results 20% [2] |
Last verified: November 2026 (exam facts on iiba.org; guide text checked section by section).
What the guide says technical visualizations are
The guide calls them "a very specific set" of visualization techniques [1]. They show insights that are statistical or mathematical in nature. Data scientists use them to find the optimal analytics approach and to confirm modelling assumptions [1].
The key contrast is with business visualizations. A business visual focuses on "one simplified message" for the business. A technical visual focuses on "discerning data patterns" [1]. Section 2.3.6 says it in one line: technical visualizations are used to analyse the data, business visualizations to interpret and report results (Guide, 2.3.6).
The analyst has a clear job here [1]:
- be comfortable with common technical visualizations;
- understand the analysis that data scientists perform;
- simplify, translate and communicate the impact of modelling assumptions or insights to stakeholders.
So I would not expect a question that asks for a silhouette calculation. I would expect questions about what a chart means for the business and who should see it.
Where technical visualizations are used in the CBDA domains
Section 3.18 has no domain-by-domain subsection. The links come from chapter 2 and from neighbouring techniques:
| Domain, task or technique | What the guide says |
|---|---|
| 2.3.6, Select Techniques for Analyze Data | Used to understand patterns and signals "during exploratory data analysis and data modelling" (Guide, 2.3.6) |
| 2.4.4, Derive Insights from Data | An error residue graph shows prediction error falling as predictors are added; it helps pick the number of variables, while a marketing stakeholder wants ad spend against revenue (Guide, 2.4.4) |
| 3.2, Business Visualizations | Business visuals communicate insights; technical visuals are aids for analytics professionals (Guide, 3.2) |
| 3.11, Exploratory Data Analysis | Contour, density and pair plots show more than two variables; lag plots, autocorrelation plots and Box-Jenkins are used for time series (Guide, 3.11) |
For the whole domain, see Analyze Data, Domain 3. The neighbour technique is Business Visualizations, and the line between the two is the most likely exam question.
The eight technical visualizations in the guide
The elements differ for each visual, so the guide shows a common set [1]. I put each one next to the business question it answers.
| Visual | What it shows (guide) [1] | Business question behind it | Outside reference |
|---|---|---|---|
| Autocorrelation plot (Box-Jenkins) | Whether outcomes correlate with their own past values; near zero means random | Is there seasonality we must plan for? | Used for randomness checks and Box-Jenkins model identification [3] |
| Box-Cox plot | Whether a skewed variable needs a transformation toward normal | Can we keep the extreme cases instead of deleting them? | The best λ is where the correlation peaks [4] |
| Contour plot | A third variable as lines or rings on a 2D map; from seismic analysis | Where is the optimum or the densest area? | — |
| Correlation map | Pairwise linear dependence from -1 to 1 | Which drivers say the same thing twice? | — |
| PCA plot | A few principal components instead of many variables | What groups exist, in plain words? | — |
| Residual plot | Patterns in model errors; inverted U means non-linearity, uneven spread means heteroskedasticity | Where is the model wrong, and for whom? | Detects non-linearity, unequal error variances and outliers [5] |
| ROC curve | True vs false positive rate across thresholds; more AUC is better | How many good customers do we bother to catch one bad one? | Top-left corner is the ideal point [8] |
| Silhouette plot | How well each cluster is separated, score from -1 to 1 | How many segments are real? | Best value 1, worst -1, near 0 means overlap [6] |
A few details from the guide are worth remembering:
- Box-Cox. Many restaurants have 1 to 20 reviews, a few have 2,000. Deleting the 2,000-review restaurants as outliers "may not be consistent with the business context", because more reviews mean more confidence [1]. The analyst proposes a transformation based on the context.
- Correlation map. In attrition analysis, years of experience and designation may both influence the result. Models that assume independent variables use the map, with p-values, to remove unwanted variables [1].
- PCA. In the guide's used car example, PC1 groups cars with a large engine, high fuel expenditure and high power, and PC2 is the opposite [1]. Analysts are often asked to interpret components "in business terms", for example as customer segments [1].
- Residuals. Heteroskedasticity can indicate subgroups in the data with different attributes, which may change the model [1].
- Silhouette. In the guide's example, 10 segments were chosen, but only four to five dominate. A narrow cluster may need splitting; clusters with negative values may be removed [1].
A worked example: translating two charts
This example is invented, and so are the numbers. A grocery chain gets two models from its data science team: a weekly sales forecast and a fraud score for online payments.
Part 1: the forecast.
- The residual plot shows a fan: small stores miss by about ±2%, large stores by about ±9%. The guide reads this pattern as heteroskedasticity and a hint of subgroups [1].
- The autocorrelation plot of residuals shows a spike at lag 52. The errors repeat every year, so the model misses yearly seasonality. Random errors would show autocorrelation near zero [1, 3].
- The analyst's translation: the forecast is fine for small stores; for large stores it misses holiday weeks, so we should model large stores separately and add the holiday calendar.
Part 2: the fraud score.
- The ROC curve looks strong. But the business does not run "the curve". It runs one threshold [1].
- Facts for one month: 50,000 payments, 1% fraud, so 500 frauds. The chosen threshold flags 2% of payments, 1,000 alerts.
- At this threshold the true positive rate is 70%: 350 frauds caught. The other 650 alerts are good customers. The false positive rate is 650 / 49,500 = 1.31%, and precision is 350 / 1,000 = 35%.
- The analyst's translation: we catch 7 frauds out of 10; out of every 100 alerts, 35 are fraud; about 650 good customers a month get an extra check. The steering committee can now argue about the threshold in its own terms.
None of these steps needed code from the analyst. Each needed one question: what does this pattern mean for a decision?
Strengths and limitations
The guide lists four strengths and three limitations [1]:
| Strengths [1] | Limitations [1] |
|---|---|
| Insights on model selection, assumptions and fit that improve models and reduce development time | Hard to interpret without the technical foundation |
| Well defined for many problem types, so a systematic approach is possible | Interpretations may bias the models, which are built only on the available data and may not match the business context |
| They show how and why modelling choices were made; without them, the rationale "can be lost" for stakeholders | They need business and domain knowledge, which is hard to turn into empirical rules |
| Many tools exist to create standard visuals |
My take: the second limitation is the most important for the analyst. A chart built on available data can look convincing and still miss the business context. That is exactly where the analyst adds value.
What research and practitioners say
The research says that plots catch what numbers hide, and that some plots mislead. Four data sets with almost the same summary statistics look completely different on a chart, which is the classic argument for plotting [10]. On the other side:
- ROC on rare events. On imbalanced data, the visual reading of a ROC plot "can be deceptive"; precision-recall plots change with the class ratio, ROC plots do not [9].
- AUC and costs. A radiologist and two medical statisticians note that ROC AUC ignores prevalence and the different costs of errors [12].
- PCA artefacts. A 2022 study in population genetics shows that PCA results can be artefacts of the data and can be steered toward a desired outcome [11]. A practitioner newsletter adds a practical check: a 2D PCA plot misleads when the first two components explain little variance, 55% in its example [14].
- Metrics alone do not decide. A study with subject-matter experts found that metrics alone were not enough for decisions, and visual comparisons that tell a story were preferred [13].
Practitioners describe the translation problem from both sides:
- A better curve for business owners. A healthcare AI CTO replaced ROC with an "outcome capture" curve. At a 10% intervention threshold, the model captured 22% of readmissions against 16% for a standard index [15]. He still keeps ROC AUC when one number is needed [15].
- One model, four audiences. A fraud data scientist restates one model four ways: AUC 0.902 for data scientists, an "insult rate" of 0.13% for customer support, 44% precision after review for trust and safety, and $29,000 of prevented monthly loss for finance. The post presents them as a worked example [16].
- High AUC, small gain. A head of data science in identity fraud shows that at 1% fraud even a weak score reaches AUC about 0.92, and that further AUC gains can come at about 2% precision. His advice is to pick a cutoff and estimate the losses [17].
- Residuals vs temperature. An energy load forecaster plotted residuals against temperature and found the model "mis-specified": it under-predicted at low and high temperatures and over-predicted in the middle [18].
- AUC 0.68 can be useful. A data scientist who worked on an early-warning system for a chronic lung disease writes that a model with AUC 0.68 was still useful, because the alternative was doing nothing [19].
- Cannot defend the curve. In an anonymous professional forum, a data scientist with AUC 0.9 could not defend the model to a manager in dollars. Peers advised showing the business what false positives and false negatives cost [20].
- Lift instead of ROC. A data scientist turned product manager shows executives a lift chart instead of a confusion matrix and ROC curve [21].
- Silhouette vs personas. An applied ML teacher describes a segmentation where the silhouette score was best at two clusters (0.41), but marketing needed four personas. His advice is to say openly that k=4 came from the business narrative [22].
From the exam side, a data analyst with more than ten years of experience said that many CBDA questions "had two confusing answers" [30]. Visualization questions look like that: two charts are correct, only one fits the audience.
Experts disagree on technical visualizations
Practitioners agree that these charts are needed inside the team. They disagree on who should see them and how far to trust them.
1. Keep ROC curves internal. A data science newsletter author recommends using the ROC curve only "internally" [23]. The lift chart author [21] takes the same line. The other side does not show raw ROC either, but translates it: the outcome capture curve [15], the four-audience fraud model [16], and the study that found visual stories work better for experts [13]. The guide is closer to the second side: the analyst simplifies and translates [1].
2. AUC as a single number. A guest author on a clinical statistics blog calls the AUROC "a reasonable initial screen", with a warning that a well-discriminating model can be poorly calibrated [24]. A data mining consultant argues that AUC "should, in fact, never be used" and that models should be compared in the operating region [25]. Even the fix is disputed: precision-recall curves are recommended for imbalanced data [9], while a 2024 paper argues they are not generally better under class imbalance [26].
3. Eyes or tests. Statisticians behind a 2023 study say conventional tests are "too sensitive" and argue for a structured visual reading of residual plots, the lineup protocol [27]. A statistics software FAQ notes that normality tests are "less useful than some people guess": they have little power in small samples, and in large samples non-normality matters less [28]. My view: for the analyst, the plot is the conversation and the test is the footnote.
4. Who sets the number of segments. The silhouette was designed to help choose an "appropriate" number of clusters [7], and it is often used when the elbow method gives no clear answer [29]. The practitioner in [22] lets the business choose and discloses it. In the guide's example, the number of segments is read from the plot; the guide does not discuss business input on it [1].
5. How much to trust a PCA picture. The guide treats PCA groups as segments the analyst can name [1]. The artefact study [11] and the variance check [14] say to name them only after checking how much variance the picture keeps.
How it differs from similar techniques
Technical visualizations are easy to confuse with three neighbours:
| Compared with | Difference in the guide |
|---|---|
| Business Visualizations (3.2) | Communicate insights to business stakeholders; technical visuals are aids for analytics professionals (Guide, 3.2) |
| Exploratory Data Analysis (3.11) | A process where visuals are the "primary vehicle"; it uses univariate, bivariate and special purpose plots (Guide, 3.11) |
| Descriptive and Inferential Statistics (3.9) | Descriptive statistics describe the data; inferential statistics support predictive and prescriptive modelling (Guide, 2.3.6). Technical visuals show the patterns behind those numbers |
For the next step, read Exploratory Data Analysis, where most of these plots first appear, and Descriptive and Inferential Statistics for the numbers behind them.
Exam traps for technical visualizations
| Scenario | Tempting answer | Guide reading |
|---|---|---|
| Show the sponsor why the model was chosen | Present the ROC curve as is | Translate the modelling choice into business impact [1] |
| Which domain lists technical visualizations? | Interpret and Report Results | Analyze Data, 2.3.6 (Guide, 2.3.6) |
| Restaurants with 2,000 reviews skew the data | Remove them as outliers | Consider a normal transformation; outliers may matter in the business context [1] |
| Residual plot with an inverted U | The model fits well | A non-linear relationship is present [1] |
| Residual spread grows with the value | Random noise | Heteroskedasticity, possibly subgroups [1] |
| Autocorrelation near zero at all lags | Strong seasonality | The data behave randomly [1] |
| Ten segments, four to five dominant | Keep all ten | Consider fewer clusters [1] |
| Compare two classifiers | Pick the one with lower AUC | More area under the curve means a better model for classification [1] |
| Who builds technical visuals? | The analyst alone | Data scientists use them; analysts understand and translate [1] |
For chart-choice practice, use the CBDA data visualization questions.
Check yourself: CBDA practice questions
1. A data scientist shows a silhouette plot: 10 clusters, but only four have wide silhouettes, and two have negative values. What should the analyst suggest? A) Keep 10 clusters. B) Reduce the number of clusters and consider removing the negative ones. C) Switch to a ROC curve. D) Remove all narrow clusters from the data set.
Answer: B. The guide's example reads the same pattern as a reason to reduce clusters and remove weak ones [1].
2. A marketing director asks how ad spend relates to revenue. The team has an error residue graph and a chart of ad spend against revenue. Which should the analyst show? A) The error residue graph. B) Both, with no explanation. C) The ad spend chart. D) Neither.
Answer: C. The guide uses this exact contrast: the residue graph helps choose variables, the marketing stakeholder wants ad spend against revenue (Guide, 2.4.4).
3. Which statement about technical visualizations matches the guide? A) They are the main tool for reporting results to executives. B) They help confirm modelling assumptions, and analysts translate them for stakeholders. C) Analysts never need to understand them. D) They replace business visualizations.
Answer: B. They confirm modelling assumptions, and analysts simplify, translate and communicate [1].
Conclusion: the gap between the guide and real model reviews
The guide's eight visuals match what data scientists use, but in practice the hard part is the translation, and experts argue about some of the charts themselves.
| Area | Guide | Reality | Delta |
|---|---|---|---|
| Audience | Mostly the data team; analysts translate [1] | Teams build business versions: outcome capture, cost per alert, lift [15, 16, 21] | No gap |
| ROC and AUC | More AUC means a better model [1] | AUC can mislead on rare events and ignores costs [9, 12, 17] | Reduced to a first screen |
| Residual reading | Patterns show fit, non-linearity, subgroups [1] | Plots caught real mis-specification; tests are argued over [18, 27, 28] | No gap |
| Number of segments | Suggested by the silhouette plot [1] | Can be set by the business narrative, which should then be disclosed [22] | Added by practitioners |
| PCA | Name components in business terms [1] | Check explained variance first; plots can be artefacts [11, 14] | Added by practitioners |
So, for the exam, learn what each of the eight visuals shows and who it is for. For real work, I would add one habit: never forward a technical chart without one sentence that says what it means for a decision.
Related topics
- CBDA exam simulator with full mock exams.
FAQ
What are technical visualizations in the CBDA guide?
They are technique 3.18 of the IIBA Guide to Business Data Analytics: charts that data scientists use to understand data, choose an analytics approach and confirm modelling assumptions. Examples are ROC curves, residual plots, PCA plots and silhouette plots.
What is the difference between technical and business visualizations?
Technical visualizations help the analytics team analyse data and focus on data patterns. Business visualizations communicate one simplified message to business stakeholders. The guide lists technical visualizations under Analyze Data and business visualizations for interpreting and reporting results.
Does a business analyst need to build ROC curves or residual plots for the CBDA?
No. The guide says data scientists use these visuals. Analysts are comfortable with the common ones, understand the analysis, and simplify, translate and communicate the impact of modelling assumptions to stakeholders.
Which technical visualizations does the guide list?
Eight common ones: autocorrelation plots, Box-Cox plots, contour plots, correlation maps, principal component analysis plots, residual plots, ROC curves and silhouette plots.
Which CBDA domain covers technical visualizations?
Section 2.3.6, Select Techniques for Analyze Data, lists them. Analyze Data is Domain 3 and makes up 16% of the CBDA exam. Section 2.4.4 in Domain 4 also uses a technical visual as an example.
Sources
- IIBA: Guide to Business Data Analytics, 3.18 Technical Visualizations
- IIBA: Certification in Business Data Analytics (CBDA)
- NIST/SEMATECH e-Handbook: Autocorrelation plot
- NIST/SEMATECH e-Handbook: Box-Cox normality plot
- Penn State STAT 462: Residuals vs fits plot
- scikit-learn: silhouette_score
- Silhouettes: a graphical aid to the interpretation and validation of cluster analysis (1987)
- scikit-learn: Receiver operating characteristic example
- PLOS ONE: The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets (2015)
- Anscombe's quartet (2024)
- Scientific Reports: PCA-based findings in population genetic studies are highly biased (2022)
- European Radiology: Disadvantages of using the area under the ROC curve to assess imaging tests (2015)
- arXiv 2205.05749: a study of model performance visuals for subject-matter experts (2022)
- Avoid using PCA for visualization (2024)
- Move over ROC curve, there's a better graph for population health (2019)
- Numbers for the whole company: the value of machine learning (2022)
- How not to evaluate fraud models (2021)
- The story is in the residuals (2019)
- Performance metrics in data science are widely misunderstood and misused
- Forum thread: manager doesn't understand data science (2023)
- Communicating the business value of data science (2016)
- Clustering isn't discovery, it's a hypothesis you still have to test (2026)
- Actually existing AI and ML (2022)
- Statistical Thinking blog: guest post on machine learning confusion (2018)
- AUC: a fatally flawed model metric
- arXiv 2401.06091: AUROC and AUPRC under class imbalance (2024)
- arXiv 2308.05964: reading residual plots, a study of the lineup protocol and conventional tests (2023)
- GraphPad FAQ: Normality tests
- Built In: the elbow method (2025)
- CBDA certification journey of a data analyst (2024)